Skip to content

GitLab

Commit 75535dd8 authored by Aleksander Dudek's avatar Aleksander Dudek
Browse files

Merge branch 'develop' into ck_tile_gemmkernel_reuse

parents b4f65acf 77a38e02
Show whitespace changes
Inline Side-by-side
Showing with 894 additions and 111 deletions
Dockerfile
View file @ 75535dd8
FROM ubuntu:20.04 FROM ubuntu:22.04
ARG DEBIAN_FRONTEND=noninteractive ARG DEBIAN_FRONTEND=noninteractive
ARG ROCMVERSION=6.3 ARG ROCMVERSION=6.3
ARG compiler_version="" ARG compiler_version=""
...@@ -48,6 +48,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -48,6 +48,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
libnuma-dev \ libnuma-dev \
libpthread-stubs0-dev \ libpthread-stubs0-dev \
llvm-amdgpu \ llvm-amdgpu \
mpich \
net-tools \ net-tools \
pkg-config \ pkg-config \
python \ python \
...@@ -70,7 +71,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -70,7 +71,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
rm -rf /var/lib/apt/lists/* && \ rm -rf /var/lib/apt/lists/* && \
rm -rf amdgpu-install* && \ rm -rf amdgpu-install* && \
# Remove unnecessary rocm components that take a lot of space # Remove unnecessary rocm components that take a lot of space
apt-get remove -y rocblas rocfft rocsparse composablekernel-dev apt-get remove -y rocblas rocfft rocsparse composablekernel-dev hipblaslt
# Update the cmake to version 3.27.5 # Update the cmake to version 3.27.5
RUN pip install --upgrade cmake==3.27.5 && \ RUN pip install --upgrade cmake==3.27.5 && \
......
Dockerfile.compiler
View file @ 75535dd8
ARG BASE_DOCKER="rocm/composable_kernel:ck_ub20.04_rocm6.3" ARG BASE_DOCKER="rocm/composable_kernel:ck_ub22.04_rocm6.3"
FROM $BASE_DOCKER FROM $BASE_DOCKER
ARG compiler_version="" ARG compiler_version=""
ARG compiler_commit="" ARG compiler_commit=""
......
Jenkinsfile
View file @ 75535dd8
...@@ -40,10 +40,10 @@ def getBaseDockerImageName(){ ...@@ -40,10 +40,10 @@ def getBaseDockerImageName(){
else{ else{
def ROCM_numeric = "${params.ROCMVERSION}" as float def ROCM_numeric = "${params.ROCMVERSION}" as float
if ( ROCM_numeric < 6.4 ){ if ( ROCM_numeric < 6.4 ){
img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}" img = "${env.CK_DOCKERHUB}:ck_ub22.04_rocm${params.ROCMVERSION}"
} }
else{ else{
img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}" img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub22.04_rocm${params.ROCMVERSION}"
} }
} }
return img return img
...@@ -357,7 +357,7 @@ def buildHipClangJob(Map conf=[:]){ ...@@ -357,7 +357,7 @@ def buildHipClangJob(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm") def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group // Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined" def dockerOpts="-u root --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) { if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 " dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
} }
...@@ -377,7 +377,7 @@ def buildHipClangJob(Map conf=[:]){ ...@@ -377,7 +377,7 @@ def buildHipClangJob(Map conf=[:]){
gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') { gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') { withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 48, unit: 'HOURS') timeout(time: 20, unit: 'HOURS')
{ {
cmake_build(conf) cmake_build(conf)
} }
...@@ -426,7 +426,7 @@ def Build_CK(Map conf=[:]){ ...@@ -426,7 +426,7 @@ def Build_CK(Map conf=[:]){
def prefixpath = conf.get("prefixpath", "/opt/rocm") def prefixpath = conf.get("prefixpath", "/opt/rocm")
// Jenkins is complaining about the render group // Jenkins is complaining about the render group
def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined" def dockerOpts="-u root --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
if (conf.get("enforce_xnack_on", false)) { if (conf.get("enforce_xnack_on", false)) {
dockerOpts = dockerOpts + " --env HSA_XNACK=1 " dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
} }
...@@ -449,7 +449,7 @@ def Build_CK(Map conf=[:]){ ...@@ -449,7 +449,7 @@ def Build_CK(Map conf=[:]){
try { try {
(retimage, image) = getDockerImage(conf) (retimage, image) = getDockerImage(conf)
withDockerContainer(image: image, args: dockerOpts) { withDockerContainer(image: image, args: dockerOpts) {
timeout(time: 5, unit: 'MINUTES'){ timeout(time: 2, unit: 'MINUTES'){
sh 'rocminfo | tee rocminfo.log' sh 'rocminfo | tee rocminfo.log'
if ( !runShell('grep -n "gfx" rocminfo.log') ){ if ( !runShell('grep -n "gfx" rocminfo.log') ){
throw new Exception ("GPU not found") throw new Exception ("GPU not found")
...@@ -465,7 +465,7 @@ def Build_CK(Map conf=[:]){ ...@@ -465,7 +465,7 @@ def Build_CK(Map conf=[:]){
throw e throw e
} }
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') { withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 12, unit: 'HOURS') timeout(time: 20, unit: 'HOURS')
{ {
//check whether to run performance tests on this node //check whether to run performance tests on this node
def arch_type = 0 def arch_type = 0
...@@ -620,7 +620,7 @@ def process_results(Map conf=[:]){ ...@@ -620,7 +620,7 @@ def process_results(Map conf=[:]){
} }
withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') { withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
timeout(time: 1, unit: 'HOURS'){ timeout(time: 15, unit: 'MINUTES'){
try{ try{
dir("script"){ dir("script"){
if (params.RUN_CK_TILE_FMHA_TESTS){ if (params.RUN_CK_TILE_FMHA_TESTS){
...@@ -675,8 +675,8 @@ def process_results(Map conf=[:]){ ...@@ -675,8 +675,8 @@ def process_results(Map conf=[:]){
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.3;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.3;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
0 21 * * * % ROCMVERSION=6.3;hipTensor_test=true;RUN_CODEGEN_TESTS=true 0 21 * * * % ROCMVERSION=6.3;hipTensor_test=true;RUN_CODEGEN_TESTS=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true 0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true 0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false 0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
0 13 * * * % BUILD_LEGACY_OS=true''' : "" 0 13 * * * % BUILD_LEGACY_OS=true''' : ""
...@@ -763,8 +763,8 @@ pipeline { ...@@ -763,8 +763,8 @@ pipeline {
description: "Test building instances for various architectures simultaneously (default: OFF)") description: "Test building instances for various architectures simultaneously (default: OFF)")
booleanParam( booleanParam(
name: "BUILD_GFX12", name: "BUILD_GFX12",
defaultValue: false, defaultValue: true,
description: "Build CK and run tests on gfx12 (default: OFF)") description: "Build CK and run tests on gfx12 (default: ON)")
booleanParam( booleanParam(
name: "NINJA_BUILD_TRACE", name: "NINJA_BUILD_TRACE",
defaultValue: false, defaultValue: false,
......
codegen/test/rtc/include/rtc/hip.hpp
View file @ 75535dd8
...@@ -4,6 +4,7 @@ ...@@ -4,6 +4,7 @@
#include <hip/hip_runtime_api.h> #include <hip/hip_runtime_api.h>
#include <memory> #include <memory>
#include <string> #include <string>
#include <stdexcept>
namespace rtc { namespace rtc {
......
example/ck_tile/01_fmha/codegen/cpp_symbol_map.py
View file @ 75535dd8
...@@ -2,10 +2,17 @@ ...@@ -2,10 +2,17 @@
# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. # Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
# generate kernel instances to speed up compilation # generate kernel instances to speed up compilation
DTYPE_MAP = { FWD_DTYPE_MAP = {
"fp16": "ck_tile::fp16_t", "fp16" : "FmhaFwdFp16",
"bf16": "ck_tile::bf16_t", "bf16" : "FmhaFwdBf16",
"fp8" : "ck_tile::fp8_t" "fp8" : "FmhaFwdFp8",
"fp8fp16": "FmhaFwdFp8Fp16",
"fp8bf16": "FmhaFwdFp8Bf16"
}
BWD_DTYPE_MAP = {
"fp16": "FmhaBwdFp16",
"bf16": "FmhaBwdBf16"
} }
MASK_IMPL = { MASK_IMPL = {
......
example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py
View file @ 75535dd8
...@@ -283,7 +283,7 @@ class FmhaBwdApiPool: ...@@ -283,7 +283,7 @@ class FmhaBwdApiPool:
inners = inners + FMHA_BWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_pipeline_enum=BWD_DQDKDV_PIPELINE_ENUM_MAP[trait.pipeline], inners = inners + FMHA_BWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_pipeline_enum=BWD_DQDKDV_PIPELINE_ENUM_MAP[trait.pipeline],
F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias],
F_bias=BIAS_MAP[trait.bias], F_dbias=BOOL_MAP[trait.dbias], F_dropout_check=DROPOUT_CHECK_MAP[trait.dropout], F_dropout=DROPOUT_MAP[trait.dropout], F_bias=BIAS_MAP[trait.bias], F_dbias=BOOL_MAP[trait.dbias], F_dropout_check=DROPOUT_CHECK_MAP[trait.dropout], F_dropout=DROPOUT_MAP[trait.dropout],
F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=DTYPE_MAP[dtype], F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=BWD_DTYPE_MAP[dtype],
F_spad0=BOOL_MAP[trait.spad], F_spad1=BOOL_MAP[spad1], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_spad0=BOOL_MAP[trait.spad], F_spad1=BOOL_MAP[spad1], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_deterministic=BOOL_MAP[trait.deterministic]) F_deterministic=BOOL_MAP[trait.deterministic])
...@@ -360,7 +360,7 @@ class FmhaBwdDQDKDVKernel: ...@@ -360,7 +360,7 @@ class FmhaBwdDQDKDVKernel:
FMHA_BWD_DQ_DK_DV_KERNEL_BODY.format( FMHA_BWD_DQ_DK_DV_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0, F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0, F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0, F_bk0 = self.F_tile.F_bk0,
...@@ -469,7 +469,7 @@ def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -469,7 +469,7 @@ def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list() gen = list()
api_pool = FmhaBwdApiPool(mask_impl) api_pool = FmhaBwdApiPool(mask_impl)
for dtype in DTYPE_MAP.keys(): for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype) d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
...@@ -585,7 +585,7 @@ class FmhaBwdOGradDotOKernel: ...@@ -585,7 +585,7 @@ class FmhaBwdOGradDotOKernel:
FMHA_BWD_DOT_DO_O_KERNEL_BODY.format( FMHA_BWD_DOT_DO_O_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_spad = BOOL_MAP[self.F_spad], F_spad = BOOL_MAP[self.F_spad],
F_dvpad = BOOL_MAP[self.F_dvpad], F_dvpad = BOOL_MAP[self.F_dvpad],
F_mode = MODE_MAP[self.F_mode], F_mode = MODE_MAP[self.F_mode],
...@@ -616,7 +616,7 @@ def get_bwd_dot_do_o_blobs() -> List[FmhaBwdOGradDotOKernel]: ...@@ -616,7 +616,7 @@ def get_bwd_dot_do_o_blobs() -> List[FmhaBwdOGradDotOKernel]:
gen = list() gen = list()
for dtype in DTYPE_MAP.keys(): for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype) d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
...@@ -716,7 +716,7 @@ class FmhaBwdConvertQGradKernel: ...@@ -716,7 +716,7 @@ class FmhaBwdConvertQGradKernel:
FMHA_BWD_CONVERT_DQ_KERNEL_BODY.format( FMHA_BWD_CONVERT_DQ_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = BWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_bm0, F_bm0 = self.F_bm0,
F_bn0 = self.F_bn0, F_bn0 = self.F_bn0,
F_spad = BOOL_MAP[self.F_spad], F_spad = BOOL_MAP[self.F_spad],
...@@ -751,7 +751,7 @@ def get_bwd_convert_dq_blobs() -> List[FmhaBwdConvertQGradKernel]: ...@@ -751,7 +751,7 @@ def get_bwd_convert_dq_blobs() -> List[FmhaBwdConvertQGradKernel]:
gen = list() gen = list()
for dtype in DTYPE_MAP.keys(): for dtype in BWD_DTYPE_MAP.keys():
d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype) d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py
View file @ 75535dd8
...@@ -282,7 +282,7 @@ class FmhaFwdApiPool: ...@@ -282,7 +282,7 @@ class FmhaFwdApiPool:
F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max, F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype]) F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if' if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners) per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if' if_i = 'if' if i == 0 else 'else if'
...@@ -339,7 +339,7 @@ class FmhaFwdKernel: ...@@ -339,7 +339,7 @@ class FmhaFwdKernel:
FMHA_FWD_KERNEL_BODY.format( FMHA_FWD_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0, F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0, F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0, F_bk0 = self.F_tile.F_bk0,
...@@ -462,6 +462,9 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm ...@@ -462,6 +462,9 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
# no need lse/dropout kernels # no need lse/dropout kernels
for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()): for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask)) pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else: else:
assert False assert False
return pipelines return pipelines
...@@ -469,7 +472,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm ...@@ -469,7 +472,7 @@ def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[Fm
gen = list() gen = list()
api_pool = FmhaFwdApiPool(mask_impl) api_pool = FmhaFwdApiPool(mask_impl)
for dtype in DTYPE_MAP.keys(): for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_tile_dict_from_dtype(dtype) d = get_fmha_fwd_tile_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_appendkv.py
View file @ 75535dd8
...@@ -181,7 +181,7 @@ class FmhaFwdAppendKVApiPool: ...@@ -181,7 +181,7 @@ class FmhaFwdAppendKVApiPool:
inners = inners + FMHA_FWD_APPENDKV_API_INNER_DISPATCH.format(F_if=if_k, F_vlayout=LAYOUT_MAP[trait.vlayout], inners = inners + FMHA_FWD_APPENDKV_API_INNER_DISPATCH.format(F_if=if_k, F_vlayout=LAYOUT_MAP[trait.vlayout],
F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_rope_check=ROPE_CHECK_MAP[trait.rope], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_rope_check=ROPE_CHECK_MAP[trait.rope],
F_pagedkv=BOOL_MAP[trait.pagedkv], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_pagedkv=BOOL_MAP[trait.pagedkv], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=DTYPE_MAP[dtype]) F_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if' if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners) per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if' if_i = 'if' if i == 0 else 'else if'
...@@ -216,7 +216,7 @@ class FmhaFwdAppendKVKernel: ...@@ -216,7 +216,7 @@ class FmhaFwdAppendKVKernel:
FMHA_FWD_APPENDKV_KERNEL_BODY.format( FMHA_FWD_APPENDKV_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bs = self.F_tile.F_bs, F_bs = self.F_tile.F_bs,
F_bsk = self.F_tile.F_bsk, F_bsk = self.F_tile.F_bsk,
F_bd = self.F_tile.F_bd, F_bd = self.F_tile.F_bd,
...@@ -301,6 +301,9 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -301,6 +301,9 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
elif dtype in ['fp8', 'bf8']: elif dtype in ['fp8', 'bf8']:
# rope/paged-kv is not supported # rope/paged-kv is not supported
pipelines.append(FmhaFwdAppendKVPipeline('col', 't', 't', 't', 't', 'no', 'f')) pipelines.append(FmhaFwdAppendKVPipeline('col', 't', 't', 't', 't', 'no', 'f'))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else: else:
assert False assert False
return pipelines return pipelines
...@@ -308,7 +311,7 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -308,7 +311,7 @@ def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list() gen = list()
api_pool = FmhaFwdAppendKVApiPool(mask_impl) api_pool = FmhaFwdAppendKVApiPool(mask_impl)
for dtype in DTYPE_MAP.keys(): for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_appendkv_tile_dict_from_dtype(dtype) d = get_fmha_fwd_appendkv_tile_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
......
example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py
View file @ 75535dd8
...@@ -435,7 +435,7 @@ class FmhaFwdSplitKVApiPool: ...@@ -435,7 +435,7 @@ class FmhaFwdSplitKVApiPool:
F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max, F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
F_hdim=hdim, F_dtype=DTYPE_MAP[dtype]) F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
if_j = 'if' if j == 0 else 'else if' if_j = 'if' if j == 0 else 'else if'
per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners) per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
if_i = 'if' if i == 0 else 'else if' if_i = 'if' if i == 0 else 'else if'
...@@ -472,7 +472,7 @@ class FmhaFwdSplitKVKernel: ...@@ -472,7 +472,7 @@ class FmhaFwdSplitKVKernel:
FMHA_FWD_SPLITKV_KERNEL_BODY.format( FMHA_FWD_SPLITKV_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0, F_bm0 = self.F_tile.F_bm0,
F_bn0 = self.F_tile.F_bn0, F_bn0 = self.F_tile.F_bn0,
F_bk0 = self.F_tile.F_bk0, F_bk0 = self.F_tile.F_bk0,
...@@ -552,7 +552,7 @@ class FmhaFwdSplitKVCombineKernel: ...@@ -552,7 +552,7 @@ class FmhaFwdSplitKVCombineKernel:
FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY.format( FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY.format(
F_idx = self.F_idx, F_idx = self.F_idx,
F_hdim = self.F_hdim, F_hdim = self.F_hdim,
F_dtype = DTYPE_MAP[self.F_dtype], F_dtype = FWD_DTYPE_MAP[self.F_dtype],
F_bm0 = self.F_tile.F_bm0, F_bm0 = self.F_tile.F_bm0,
F_bn1 = self.F_tile.F_bn1, F_bn1 = self.F_tile.F_bn1,
F_spad = BOOL_MAP[self.F_pipeline.F_spad], F_spad = BOOL_MAP[self.F_pipeline.F_spad],
...@@ -644,6 +644,9 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -644,6 +644,9 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
elif dtype in ['fp8', 'bf8']: elif dtype in ['fp8', 'bf8']:
for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()): for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 't', squant, 'f', mask)) pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 't', squant, 'f', mask))
elif dtype in ['fp8fp16', 'fp8bf16']:
# TODO
None
else: else:
assert False assert False
return pipelines return pipelines
...@@ -651,7 +654,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> ...@@ -651,7 +654,7 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
gen = list() gen = list()
api_pool = FmhaFwdSplitKVApiPool(mask_impl) api_pool = FmhaFwdSplitKVApiPool(mask_impl)
for dtype in DTYPE_MAP.keys(): for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_tile_dict_from_dtype(dtype) d = get_fmha_fwd_tile_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
...@@ -711,7 +714,7 @@ def get_fwd_splitkv_combine_blobs(kernel_filter : Optional[str], receipt) -> Lis ...@@ -711,7 +714,7 @@ def get_fwd_splitkv_combine_blobs(kernel_filter : Optional[str], receipt) -> Lis
gen = list() gen = list()
for dtype in DTYPE_MAP.keys(): for dtype in FWD_DTYPE_MAP.keys():
d = get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype) d = get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype)
if d == None: if d == None:
continue continue
......
example/ck_tile/01_fmha/fmha_bwd.cpp
View file @ 75535dd8
...@@ -101,7 +101,7 @@ auto create_args(int argc, char* argv[]) ...@@ -101,7 +101,7 @@ auto create_args(int argc, char* argv[])
} }
// different threshold for different dtype // different threshold for different dtype
template <typename DataType> template <typename DataTypeConfig>
auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/) auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
{ {
double rtol = 1e-2; double rtol = 1e-2;
...@@ -110,7 +110,7 @@ auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/) ...@@ -110,7 +110,7 @@ auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
} }
template <> template <>
auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v) auto get_elimit<FmhaBwdBf16>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v)
{ {
double rtol = 1e-2; double rtol = 1e-2;
double atol = 1e-2; double atol = 1e-2;
...@@ -122,7 +122,7 @@ auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_ ...@@ -122,7 +122,7 @@ auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_
return ck_tile::make_tuple(rtol, atol); return ck_tile::make_tuple(rtol, atol);
} }
template <typename DataType> template <typename DataTypeConfig>
bool run(const ck_tile::ArgParser& arg_parser) bool run(const ck_tile::ArgParser& arg_parser)
{ {
std::string data_type = arg_parser.get_str("prec"); std::string data_type = arg_parser.get_str("prec");
...@@ -209,7 +209,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -209,7 +209,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q); const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q);
const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k); const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k);
using TypeConfig = FmhaBwdTypeConfig<DataType>; using TypeConfig = FmhaBwdTypeConfig<DataTypeConfig>;
using QDataType = typename TypeConfig::QDataType; using QDataType = typename TypeConfig::QDataType;
using KDataType = typename TypeConfig::KDataType; using KDataType = typename TypeConfig::KDataType;
...@@ -933,7 +933,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -933,7 +933,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
} }
// clang-format on // clang-format on
auto [rtol, atol] = get_elimit<DataType>(hdim_q, hdim_v); auto [rtol, atol] = get_elimit<DataTypeConfig>(hdim_q, hdim_v);
bool dq_cur_pass = ck_tile::check_err(dq_host_result, bool dq_cur_pass = ck_tile::check_err(dq_host_result,
dq_host_ref, dq_host_ref,
std::string("Error: QGrad Incorrect results!"), std::string("Error: QGrad Incorrect results!"),
...@@ -986,11 +986,11 @@ int main(int argc, char* argv[]) ...@@ -986,11 +986,11 @@ int main(int argc, char* argv[])
const std::string data_type = arg_parser.get_str("prec"); const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16") if(data_type == "fp16")
{ {
return run<ck_tile::half_t>(arg_parser) ? 0 : -2; return run<FmhaBwdFp16>(arg_parser) ? 0 : -2;
} }
else if(data_type == "bf16") else if(data_type == "bf16")
{ {
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2; return run<FmhaBwdBf16>(arg_parser) ? 0 : -2;
} }
return -3; return -3;
......
example/ck_tile/01_fmha/fmha_bwd.hpp
View file @ 75535dd8
...@@ -14,11 +14,19 @@ ...@@ -14,11 +14,19 @@
#include <utility> #include <utility>
#include <variant> #include <variant>
struct FmhaBwdFp16
{
};
struct FmhaBwdBf16
{
};
template <typename DataType> template <typename DataType>
struct FmhaBwdTypeConfig; struct FmhaBwdTypeConfig;
template <> template <>
struct FmhaBwdTypeConfig<ck_tile::half_t> struct FmhaBwdTypeConfig<FmhaBwdFp16>
{ {
using QDataType = ck_tile::half_t; using QDataType = ck_tile::half_t;
using KDataType = ck_tile::half_t; using KDataType = ck_tile::half_t;
...@@ -38,7 +46,7 @@ struct FmhaBwdTypeConfig<ck_tile::half_t> ...@@ -38,7 +46,7 @@ struct FmhaBwdTypeConfig<ck_tile::half_t>
}; };
template <> template <>
struct FmhaBwdTypeConfig<ck_tile::bf16_t> struct FmhaBwdTypeConfig<FmhaBwdBf16>
{ {
using QDataType = ck_tile::bf16_t; using QDataType = ck_tile::bf16_t;
using KDataType = ck_tile::bf16_t; using KDataType = ck_tile::bf16_t;
......
example/ck_tile/01_fmha/fmha_fwd.cpp
View file @ 75535dd8
...@@ -3,6 +3,7 @@ ...@@ -3,6 +3,7 @@
#include "fmha_fwd.hpp" #include "fmha_fwd.hpp"
#include "ck_tile/host.hpp" #include "ck_tile/host.hpp"
#include "ck_tile/ref/naive_attention.hpp"
#include "mask.hpp" #include "mask.hpp"
#include "rotary.hpp" #include "rotary.hpp"
#include "utils.hpp" #include "utils.hpp"
...@@ -41,7 +42,7 @@ std::ostream& operator<<(std::ostream& os, const std::vector<T>& v) ...@@ -41,7 +42,7 @@ std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("v", "1", "weather do CPU validation or not") arg_parser.insert("v", "1", "0:no validation, 2:cpu validation, 2:gpu validation(experimental)")
.insert("mode", "0", "kernel mode. 0:batch, 1:group") .insert("mode", "0", "kernel mode. 0:batch, 1:group")
.insert("b", "2", "batch size") .insert("b", "2", "batch size")
.insert("h", "8", "num of head, for q") .insert("h", "8", "num of head, for q")
...@@ -142,7 +143,7 @@ auto create_args(int argc, char* argv[]) ...@@ -142,7 +143,7 @@ auto create_args(int argc, char* argv[])
} }
// different threshold for different dtype // different threshold for different dtype
template <typename DataType> template <typename DataTypeConfig>
auto get_elimit(std::string /*init_method*/) auto get_elimit(std::string /*init_method*/)
{ {
double rtol = 1e-3; double rtol = 1e-3;
...@@ -151,7 +152,7 @@ auto get_elimit(std::string /*init_method*/) ...@@ -151,7 +152,7 @@ auto get_elimit(std::string /*init_method*/)
} }
template <> template <>
auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/) auto get_elimit<FmhaFwdBf16>(std::string /*init_method*/)
{ {
double rtol = 1e-2; double rtol = 1e-2;
double atol = 1e-2; double atol = 1e-2;
...@@ -159,7 +160,7 @@ auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/) ...@@ -159,7 +160,7 @@ auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/)
} }
template <> template <>
auto get_elimit<ck_tile::fp8_t>(std::string init_method) auto get_elimit<FmhaFwdFp8>(std::string init_method)
{ {
if(init_method == "ui" || init_method == "ni") if(init_method == "ui" || init_method == "ni")
{ {
...@@ -261,7 +262,7 @@ int override_num_splits_if_necessary( ...@@ -261,7 +262,7 @@ int override_num_splits_if_necessary(
return num_splits; return num_splits;
} }
template <typename DataType> template <typename DataTypeConfig>
bool run(const ck_tile::ArgParser& arg_parser) bool run(const ck_tile::ArgParser& arg_parser)
{ {
std::string data_type = arg_parser.get_str("prec"); std::string data_type = arg_parser.get_str("prec");
...@@ -305,8 +306,8 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -305,8 +306,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
} }
ck_tile::index_t rotary_dim = arg_parser.get_int("rotary_dim"); ck_tile::index_t rotary_dim = arg_parser.get_int("rotary_dim");
if constexpr(!(std::is_same_v<DataType, ck_tile::fp16_t> || if constexpr(!(std::is_same_v<DataTypeConfig, FmhaFwdFp16> ||
std::is_same_v<DataType, ck_tile::bf16_t>)) std::is_same_v<DataTypeConfig, FmhaFwdBf16>))
{ {
if(0 < rotary_dim) if(0 < rotary_dim)
{ {
...@@ -428,25 +429,6 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -428,25 +429,6 @@ bool run(const ck_tile::ArgParser& arg_parser)
return atoi(squant_str.c_str()) != 0 ? true : false; return atoi(squant_str.c_str()) != 0 ? true : false;
}(); }();
float range_q = arg_parser.get_float("range_q");
float range_k = arg_parser.get_float("range_k");
float range_v = arg_parser.get_float("range_v");
float range_p = arg_parser.get_float("range_p");
float range_o = arg_parser.get_float("range_o");
float dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<DataType>::max());
float scale_p = 1.f;
float scale_o = 1.f;
if(squant)
{
scale_s = scale_s * (range_q / dtype_max) * (range_k / dtype_max);
scale_p = dtype_max / range_p;
// scale_p = [max(fp8_t)/range_o] * [range_p/max(fp8_t)] * [range_v/max(fp8_t)]
scale_o = range_p * range_v / range_o / dtype_max;
}
std::string vlayout = arg_parser.get_str("vlayout"); std::string vlayout = arg_parser.get_str("vlayout");
bool lse = arg_parser.get_bool("lse"); bool lse = arg_parser.get_bool("lse");
...@@ -466,7 +448,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -466,7 +448,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
} }
bool s_randval = false; bool s_randval = false;
if(p_drop > 0.0f && do_validation) if(p_drop > 0.0f && do_validation != 0)
{ {
s_randval = true; s_randval = true;
} }
...@@ -499,7 +481,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -499,7 +481,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
const auto seqstart_k_host = to_seqstarts(seqlen_ks); const auto seqstart_k_host = to_seqstarts(seqlen_ks);
const auto seqstart_k_with_padding_host = to_seqstarts(seqlen_kpads); const auto seqstart_k_with_padding_host = to_seqstarts(seqlen_kpads);
using TypeConfig = FmhaFwdTypeConfig<DataType>; using TypeConfig = FmhaFwdTypeConfig<DataTypeConfig>;
using QDataType = typename TypeConfig::QDataType; using QDataType = typename TypeConfig::QDataType;
using KDataType = typename TypeConfig::KDataType; using KDataType = typename TypeConfig::KDataType;
...@@ -513,6 +495,28 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -513,6 +495,28 @@ bool run(const ck_tile::ArgParser& arg_parser)
using OaccDataType = typename TypeConfig::OaccDataType; using OaccDataType = typename TypeConfig::OaccDataType;
using ODataType = typename TypeConfig::ODataType; using ODataType = typename TypeConfig::ODataType;
float range_q = arg_parser.get_float("range_q");
float range_k = arg_parser.get_float("range_k");
float range_v = arg_parser.get_float("range_v");
float range_p = arg_parser.get_float("range_p");
float range_o = arg_parser.get_float("range_o");
float q_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<QDataType>::max());
float k_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<KDataType>::max());
float v_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<VDataType>::max());
float p_dtype_max = v_dtype_max; // assume p and v is the same type
float o_dtype_max = ck_tile::type_convert<float>(ck_tile::numeric<ODataType>::max());
float scale_p = 1.f;
float scale_o = 1.f;
if(squant)
{
scale_s = scale_s * (range_q / q_dtype_max) * (range_k / k_dtype_max);
scale_p = p_dtype_max / range_p;
scale_o = (o_dtype_max / range_o) * (range_p / p_dtype_max) * (range_v / v_dtype_max);
}
// accumulation numbers for performance evaluation // accumulation numbers for performance evaluation
std::size_t flop = 0, num_byte = 0; std::size_t flop = 0, num_byte = 0;
auto max_seqlen_q = auto max_seqlen_q =
...@@ -709,14 +713,14 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -709,14 +713,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
else if(init_method == "ufq" || init_method == "uf:q" || else if(init_method == "ufq" || init_method == "uf:q" ||
init_method == "3") // suitable for fp8 quantization init_method == "3") // suitable for fp8 quantization
{ {
ck_tile::FillUniformDistribution<QDataType>{-dtype_max, dtype_max, seed}(q_host); ck_tile::FillUniformDistribution<QDataType>{-q_dtype_max, q_dtype_max, seed}(q_host);
ck_tile::FillUniformDistribution<KDataType>{-dtype_max, dtype_max, seed}(k_host); ck_tile::FillUniformDistribution<KDataType>{-k_dtype_max, k_dtype_max, seed}(k_host);
ck_tile::FillUniformDistribution<KDataType>{-dtype_max, dtype_max, seed}(knew_host); ck_tile::FillUniformDistribution<KDataType>{-k_dtype_max, k_dtype_max, seed}(knew_host);
ck_tile::FillUniformDistribution<VDataType>{-dtype_max, dtype_max, seed}(v_host); ck_tile::FillUniformDistribution<VDataType>{-v_dtype_max, v_dtype_max, seed}(v_host);
ck_tile::FillUniformDistribution<VDataType>{-dtype_max, dtype_max, seed}(vnew_host); ck_tile::FillUniformDistribution<VDataType>{-v_dtype_max, v_dtype_max, seed}(vnew_host);
// bias_fp8 = qscale_bias * bias_fp32 // bias_fp8 = qscale_bias * bias_fp32
float qscale_bias = (dtype_max / range_q) * (dtype_max / range_k); float qscale_bias = (q_dtype_max / range_q) * (k_dtype_max / range_k);
// Assume bias is in [-1.f, 1.f] in original fp32 // Assume bias is in [-1.f, 1.f] in original fp32
ck_tile::FillUniformDistribution<BiasDataType>{-qscale_bias, qscale_bias, seed}(bias_host); ck_tile::FillUniformDistribution<BiasDataType>{-qscale_bias, qscale_bias, seed}(bias_host);
} }
...@@ -1118,25 +1122,75 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -1118,25 +1122,75 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< std::setprecision(2) << tflops << " TFlops, " << std::setprecision(2) << gb_per_sec << std::setprecision(2) << tflops << " TFlops, " << std::setprecision(2) << gb_per_sec
<< " GB/s" << std::flush; << " GB/s" << std::flush;
if(!do_validation) if(do_validation == 0)
{ {
std::cout << std::flush << std::endl; std::cout << std::flush << std::endl;
return true; return true;
} }
if(do_validation == 2)
{
// NOTE: use gpu to do validation
ck_tile::naive_attention_fwd_traits naive_t;
naive_t.q_type = data_type;
naive_t.k_type = data_type;
naive_t.v_type = data_type;
naive_t.o_type = data_type;
naive_t.q_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.k_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.v_layout = i_perm == 1 ? "bhsd" : "bshd";
naive_t.o_layout = o_perm == 1 ? "bhsd" : "bshd";
naive_t.variation = 0; // TODO?
ck_tile::DeviceMem o_naive_buf(o_host.get_element_space_size_in_bytes());
ck_tile::naive_attention_fwd_args naive_a;
naive_a.q_ptr = q_buf.GetDeviceBuffer();
naive_a.k_ptr = k_buf.GetDeviceBuffer();
naive_a.v_ptr = v_buf.GetDeviceBuffer();
naive_a.o_ptr = o_naive_buf.GetDeviceBuffer();
naive_a.scale_s = scale_s;
naive_a.context_len_ptr = nullptr; // used when seqlen kv come from a pointer
naive_a.page_table_ptr =
nullptr; // [batch, num_blocks] seqlen_kv is in different block(paged attn)
naive_a.hdim = hdim_q;
naive_a.hdim_v = hdim_v; // could be cross-attn, where V and Q/K hdim are different
naive_a.batch_q = batch;
naive_a.batch_kv = batch;
naive_a.batch_ratio_kv = 1; // batch_q / batch_kv
naive_a.seqlen_q = seqlen_qs[0];
naive_a.seqlen_kv = seqlen_ks[0]; // if context_len_ptr is not nullptr, ignore this field
naive_a.nhead_q = nhead;
naive_a.nhead_kv = nhead_k;
naive_a.nhead_ratio_kv = naive_a.nhead_q / naive_a.nhead_kv; // nhead_q / nhead_kv
naive_a.page_size = 0; // if paged, the seqlen-kv for each block
ck_tile::stream_config naive_s{};
naive_attention_fwd(naive_t, naive_a, naive_s);
auto o_naive_ref = o_naive_buf.ToHost<ODataType>();
o_buf.FromDevice(o_host.data()); // TODO: ugly
auto [rtol_, atol_] = get_elimit<DataTypeConfig>(init_method);
bool pass_ = ck_tile::check_err(
o_host, o_naive_ref, std::string("OUT Error: Incorrect results!"), rtol_, atol_);
std::cout << ", valid:" << (pass_ ? "y" : "n") << std::flush << std::endl;
return pass_;
}
o_buf.FromDevice(o_host.data()); o_buf.FromDevice(o_host.data());
lse_buf.FromDevice(lse_host.data()); lse_buf.FromDevice(lse_host.data());
randval_buf.FromDevice(randval_host.data()); randval_buf.FromDevice(randval_host.data());
auto p_compute_element_func = [&]() { auto p_compute_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>) if constexpr(std::is_same_v<DataTypeConfig, ck_tile::fp8_t>)
return ck_tile::scales{scale_p}; return ck_tile::scales{scale_p};
else else
return ck_tile::identity{}; return ck_tile::identity{};
}(); }();
auto oacc_element_func = [&]() { auto oacc_element_func = [&]() {
if constexpr(std::is_same_v<DataType, ck_tile::fp8_t>) if constexpr(std::is_same_v<DataTypeConfig, ck_tile::fp8_t>)
return ck_tile::composes(ck_tile::saturates<ck_tile::fp8_t>{}, return ck_tile::composes(ck_tile::saturates<ck_tile::fp8_t>{},
ck_tile::scales{scale_o}); ck_tile::scales{scale_o});
else else
...@@ -1458,7 +1512,7 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -1458,7 +1512,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
else o_host_result.ForEach([&](auto& self, auto idx) { self(idx) = o_host(b_idx, idx[1] + query_offset, idx[0], idx[2]); }); else o_host_result.ForEach([&](auto& self, auto idx) { self(idx) = o_host(b_idx, idx[1] + query_offset, idx[0], idx[2]); });
// clang-format on // clang-format on
auto [rtol, atol] = get_elimit<DataType>(init_method); auto [rtol, atol] = get_elimit<DataTypeConfig>(init_method);
bool cur_pass = ck_tile::check_err( bool cur_pass = ck_tile::check_err(
o_host_result, o_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol); o_host_result, o_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
pass &= cur_pass; pass &= cur_pass;
...@@ -1515,15 +1569,15 @@ int main(int argc, char* argv[]) ...@@ -1515,15 +1569,15 @@ int main(int argc, char* argv[])
const std::string data_type = arg_parser.get_str("prec"); const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16") if(data_type == "fp16")
{ {
return run<ck_tile::half_t>(arg_parser) ? 0 : -2; return run<FmhaFwdFp16>(arg_parser) ? 0 : -2;
} }
else if(data_type == "bf16") else if(data_type == "bf16")
{ {
return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2; return run<FmhaFwdBf16>(arg_parser) ? 0 : -2;
} }
else if(data_type == "fp8") else if(data_type == "fp8")
{ {
return run<ck_tile::fp8_t>(arg_parser) ? 0 : -2; return run<FmhaFwdFp8>(arg_parser) ? 0 : -2;
} }
return -3; return -3;
......
example/ck_tile/01_fmha/fmha_fwd.hpp
View file @ 75535dd8
...@@ -16,11 +16,35 @@ ...@@ -16,11 +16,35 @@
#include <utility> #include <utility>
#include <variant> #include <variant>
struct FmhaFwdFp16
{
};
struct FmhaFwdBf16
{
};
struct FmhaFwdFp8
{
};
struct FmhaFwdBf8
{
};
struct FmhaFwdFp8Fp16
{
};
struct FmhaFwdFp8Bf16
{
};
template <typename DataType> template <typename DataType>
struct FmhaFwdTypeConfig; struct FmhaFwdTypeConfig;
template <> template <>
struct FmhaFwdTypeConfig<ck_tile::half_t> struct FmhaFwdTypeConfig<FmhaFwdFp16>
{ {
using QDataType = ck_tile::half_t; using QDataType = ck_tile::half_t;
using KDataType = ck_tile::half_t; using KDataType = ck_tile::half_t;
...@@ -36,7 +60,7 @@ struct FmhaFwdTypeConfig<ck_tile::half_t> ...@@ -36,7 +60,7 @@ struct FmhaFwdTypeConfig<ck_tile::half_t>
}; };
template <> template <>
struct FmhaFwdTypeConfig<ck_tile::bf16_t> struct FmhaFwdTypeConfig<FmhaFwdBf16>
{ {
using QDataType = ck_tile::bf16_t; using QDataType = ck_tile::bf16_t;
using KDataType = ck_tile::bf16_t; using KDataType = ck_tile::bf16_t;
...@@ -52,7 +76,7 @@ struct FmhaFwdTypeConfig<ck_tile::bf16_t> ...@@ -52,7 +76,7 @@ struct FmhaFwdTypeConfig<ck_tile::bf16_t>
}; };
template <> template <>
struct FmhaFwdTypeConfig<ck_tile::fp8_t> struct FmhaFwdTypeConfig<FmhaFwdFp8>
{ {
using QDataType = ck_tile::fp8_t; using QDataType = ck_tile::fp8_t;
using KDataType = ck_tile::fp8_t; using KDataType = ck_tile::fp8_t;
...@@ -68,7 +92,7 @@ struct FmhaFwdTypeConfig<ck_tile::fp8_t> ...@@ -68,7 +92,7 @@ struct FmhaFwdTypeConfig<ck_tile::fp8_t>
}; };
template <> template <>
struct FmhaFwdTypeConfig<ck_tile::bf8_t> struct FmhaFwdTypeConfig<FmhaFwdBf8>
{ {
using QDataType = ck_tile::bf8_t; using QDataType = ck_tile::bf8_t;
using KDataType = ck_tile::bf8_t; using KDataType = ck_tile::bf8_t;
......
include/ck/utility/math_v2.hpp
View file @ 75535dd8
...@@ -611,7 +611,7 @@ inline __device__ int8_t neg<int8_t>(int8_t x) ...@@ -611,7 +611,7 @@ inline __device__ int8_t neg<int8_t>(int8_t x)
template <> template <>
inline __device__ half_t neg<half_t>(half_t x) inline __device__ half_t neg<half_t>(half_t x)
{ {
return __hneg(x); return __hneg(static_cast<__half>(x));
}; };
template <typename T> template <typename T>
......
include/ck_tile/README.md
View file @ 75535dd8
...@@ -45,5 +45,8 @@ our implementation of different device operators. ...@@ -45,5 +45,8 @@ our implementation of different device operators.
**[ops/epilogue]** **[ops/epilogue]**
epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues. epilogue part of our kernel. We may extend this epilogue part to let users to build their own cutomized epilogues.
**[ref]**
reference implementation of cpu or gpu. This folder is supposed to include a specific header on demand.
## examples ## examples
currently we put all ck_tile related example under [/example/ck_tile](/example/ck_tile/) folder. Please check each example's subfolder. currently we put all ck_tile related example under [/example/ck_tile](/example/ck_tile/) folder. Please check each example's subfolder.
include/ck_tile/core.hpp
View file @ 75535dd8
...@@ -54,6 +54,7 @@ ...@@ -54,6 +54,7 @@
#include "ck_tile/core/tensor/tile_window_linear.hpp" #include "ck_tile/core/tensor/tile_window_linear.hpp"
#include "ck_tile/core/tensor/tile_window_utils.hpp" #include "ck_tile/core/tensor/tile_window_utils.hpp"
#include "ck_tile/core/tensor/update_tile.hpp" #include "ck_tile/core/tensor/update_tile.hpp"
#include "ck_tile/core/utility/amd_address_space.hpp"
#include "ck_tile/core/utility/bit_cast.hpp" #include "ck_tile/core/utility/bit_cast.hpp"
#include "ck_tile/core/utility/functional.hpp" #include "ck_tile/core/utility/functional.hpp"
#include "ck_tile/core/utility/functional_with_tuple.hpp" #include "ck_tile/core/utility/functional_with_tuple.hpp"
......
include/ck_tile/ops/gemm.hpp
View file @ 75535dd8
...@@ -23,10 +23,10 @@ ...@@ -23,10 +23,10 @@
#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp" #include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp"
#include "ck_tile/ops/gemm/block/block_gemm_problem.hpp" #include "ck_tile/ops/gemm/block/block_gemm_problem.hpp"
#include "ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp" #include "ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp"
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
#include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp" #include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
#include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp" #include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
#include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp" #include "ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp"
#include "ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp" #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp"
......
include/ck_tile/ref/README.md 0 → 100644
View file @ 75535dd8
# reference
this folder contains reference implementation of a specific op. Note by including a specific header, you are including the implementation(expecially the gpu implementation) into your source code, and compile that kernel into the fatbin, hence may increase your kernel obj code length. Usually the header starts with `reference_` is a cpu reference implementation. The header starts with `naive_` contains a gpu implementation with a small launcher.
TODO: move `host/reference` under this folder
include/ck_tile/ref/naive_attention.hpp 0 → 100644
View file @ 75535dd8
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/host/host_tensor.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include <thread>
#include <string>
namespace ck_tile {
enum class naive_attention_layout_enum
{
BSHD, // [batch, seqlen, nhead, hdim]
BHSD, // [batch, nhead, seqlen, hdim]
BS3HD, // [batch, nhead, 3, seqlen, hdim], used when qkv are packed
PHSD, // [pages, nhead, page_size, hdim]
// PHSDX, // [pages, nhead, page_size/x, hdim, x], where <# used pages>*page_size = seqlen
PHDSX, // [pages, nhead, hdim/x, page_size, x], where <# used pages>*page_size = seqlen
PHDS, // [pages, nhead, hdim, page_size], where <# used pages>*page_size = seqlen
};
// will used to specialize kernel variation
enum class naive_attention_variation_enum
{
FLASH_BATCHED = 0, // standard flash attention, or xformer/sdpa, used for training
FLASH_GROUPED,
DECODE_PAGED, // decode attn, where kv token from another buffer called kvcache
};
// TODO: for simplicity, this will be used as host/device arg
struct naive_attention_fwd_args
{
void* q_ptr;
void* k_ptr;
void* v_ptr;
void* o_ptr;
void* context_len_ptr; // [batch] used when seqlen kv come from a pointer(each element is a
// number, not cumsum)
void* page_table_ptr; // [batch, max_pages_per_seq] seqlen_kv is in different block(paged attn)
void* kvscale_ptr; // [nhead, 2(kv), hdim] used for kvcache dequant
float scale_s;
int hdim;
int hdim_v; // could be cross-attn, where V and Q/K hdim are different
int batch_q;
int batch_kv;
int batch_ratio_kv; // batch_q / batch_kv
int seqlen_q; // in decode case, this should be 1
int seqlen_kv; // if context_len_ptr is not nullptr, ignore this field
int nhead_q;
int nhead_kv;
int nhead_ratio_kv; // nhead_q / nhead_kv
int page_size; // if paged, the seqlen-kv per each block
int max_pages_per_seq;
};
// this is trait for host API
struct naive_attention_fwd_traits
{
std::string q_type;
std::string k_type;
std::string v_type;
std::string o_type;
std::string q_layout;
std::string k_layout;
std::string v_layout;
std::string o_layout;
int variation; // sync with naive_attention_variation_enum
};
// this is trait for kernel template
template <naive_attention_variation_enum variation_>
struct naive_attention_fwd_kernel_traits
{
static constexpr naive_attention_variation_enum variation = variation_;
};
// for simplicity, please do not use const-reference type for the template type
template <typename QType,
typename KType,
typename VType,
typename OType,
typename AccType,
naive_attention_layout_enum QLayout,
naive_attention_layout_enum KLayout,
naive_attention_layout_enum VLayout,
naive_attention_layout_enum OLayout,
typename Traits>
struct naive_attention_fwd_kernel
{
static constexpr bool is_kvcache_i8 =
std::is_same_v<KType, int8_t> && std::is_same_v<VType, int8_t> && sizeof(QType) != 1;
// kvcache-i8 will have per head scale, we apply this scale to Q/P matrix instead of original
// K/V matrix. This can speed up conversion since Q/P usually is fp16/bf16/fp32
static constexpr bool is_kvcache_i8_forward_quant = is_kvcache_i8;
// TODO: hardcode
using KVScaleType = float;
using SoftmaxType = float;
using PType = VType; // src A of gemm2, same type as V
using p_vec_type = ext_vector_t<PType, 16 / sizeof(PType)>;
static constexpr int p_vec_elem = vector_traits<p_vec_type>::vector_size;
__host__ __device__ naive_attention_fwd_kernel() {}
template <typename T, naive_attention_layout_enum Layout>
struct addresser
{
int b, s, h, d; // batch, seqlen, nhead, hdim
T* base_ptr;
__device__ addresser(int b_, int s_, int h_, int d_, void* base_ptr_)
: b(b_), s(s_), h(h_), d(d_), base_ptr(reinterpret_cast<T*>(base_ptr_))
{
}
// TODO: all the batch/nhead offset will accumulate to the base pointer
__device__ T* get_base(int i_b, int i_h)
{
if constexpr(Layout == naive_attention_layout_enum::BSHD)
return base_ptr + i_b * s * h * d + i_h * d;
else if constexpr(Layout == naive_attention_layout_enum::BHSD)
return base_ptr + i_b * s * h * d + i_h * s * d;
}
__device__ int get_offset(int i_s, int i_d)
{
if constexpr(Layout == naive_attention_layout_enum::BSHD)
return i_s * h * d + i_d;
else if constexpr(Layout == naive_attention_layout_enum::BHSD)
return i_s * d + i_d;
}
// below set of API will directly use pointer inside this struct
__device__ void init(int i_b, int i_h) { base_ptr = get_base(i_b, i_h); }
__device__ T load(int i_s, int i_d) { return base_ptr[get_offset(i_s, i_d)]; }
__device__ void store(T value, int i_s, int i_d) { base_ptr[get_offset(i_s, i_d)] = value; }
};
template <typename T, naive_attention_layout_enum Layout>
struct page_addresser
{
int s, h, d; // page_size, nhead, hdim
static constexpr int x = 16 / sizeof(T); // pack 4 dword
T* base_ptr;
int* page_table_ptr; // TODO: page table always int
int i_h; // store current head
__device__ page_addresser(int s_, int h_, int d_, void* base_ptr_, void* pptr_)
: s(s_),
h(h_),
d(d_),
base_ptr(reinterpret_cast<T*>(base_ptr_)),
page_table_ptr(reinterpret_cast<int*>(pptr_))
{
}
__device__ int64_t get_phy_page_idx(int i_s)
{
// dynamic compute page idx is simple but slow
int page_idx = i_s / s;
int phy = page_table_ptr[page_idx];
return static_cast<int64_t>(phy);
}
__device__ int get_phy_page_offset(int i_s)
{
// dynamic compute page idx is simple but slow
return i_s % s;
}
__device__ int64_t get_offset(int i_s, int i_d)
{
int page_offset = get_phy_page_offset(i_s);
int64_t page_idx = get_phy_page_idx(i_s);
int64_t base_ = page_idx * h * s * d;
if constexpr(Layout == naive_attention_layout_enum::PHSD)
return static_cast<int64_t>(i_h * s * d + page_offset * d + i_d) + base_;
else if constexpr(Layout == naive_attention_layout_enum::PHDSX)
{
int d_r = i_d / x;
int d_x = i_d % x;
return static_cast<int64_t>(i_h * d * s + d_r * s * x + page_offset * x + d_x) +
base_;
}
else if constexpr(Layout == naive_attention_layout_enum::PHDS)
{
return static_cast<int64_t>(i_h * d * s + i_d * s + page_offset) + base_;
}
}
// below set of API will directly use pointer inside this struct
__device__ void init(int /*i_b*/, int i_h_) { i_h = i_h_; }
__device__ T load(int i_s, int i_d) { return base_ptr[get_offset(i_s, i_d)]; }
__device__ void store(T /*value*/, int /*i_s*/, int /*i_d*/) {}
};
template <typename T>
struct kvscale_addresser
{
int h, d; // nhead, hdim
T* base_ptr;
__device__ kvscale_addresser(int h_, int d_, void* p_)
: h(h_), d(d_), base_ptr(reinterpret_cast<T*>(p_))
{
}
__device__ int get_offset(int i_h, int i_d, int i_kv /*0 or 1*/)
{
// [h, 2, d]
return i_h * 2 * d + i_kv * d + i_d;
}
__device__ T load(int i_h, int i_d, int i_kv)
{
return base_ptr[get_offset(i_h, i_d, i_kv)];
}
};
__device__ __host__ static constexpr int get_block_size() { return 256; }
// for simpliciy, 1 WG always compute 1 token along q, compute all token along kv
// compute all hdim from q, compute WG_SIZE hdim from v
// 1) in prefill case, seqlen_q >= 1, seqlen_kv >= 1, batch_q=batch_kv
// 2) in decode case, seqlen_q = 1, batch_q is input num-tokens, batch_kv is 1
// 3) in paged-attn case, we still use 1 WG compute all the seqlen-kv for simplicity
// TODO: could support split-kv to validate intermediate logsum
__host__ static dim3 get_grid_size(naive_attention_fwd_args args)
{
constexpr int wg_size = get_block_size();
auto g =
dim3((args.hdim_v + wg_size - 1) / wg_size, args.seqlen_q, args.batch_q * args.nhead_q);
return g;
}
// reduce single pixel within a wave
template <typename T, typename F>
__device__ constexpr T wave_reduce(T local, F reduce_f)
{
// constexpr int wave_size = 64;
constexpr int reduce_stage = 6; // 1<<6=64
T v_local = local;
#pragma unroll
for(int i_stage = 0; i_stage < reduce_stage; i_stage++)
{
int src_lane = __lane_id() ^ (1 << i_stage);
int32_t v_remote_tmp =
__builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(v_local));
T v_remote = bit_cast<T>(v_remote_tmp);
v_local = reduce_f(v_local, v_remote);
}
return v_local;
}
// Note: this function must be called after wave_reduce
// Note: better not use this under if...else... with thread divergence (syncthreads)
template <typename T, typename F>
__device__ constexpr T cross_wave_reduce(T local, F reduce_f, T* smem)
{
constexpr int waves = 4;
constexpr int wave_size = 64;
int lane_id = threadIdx.x % wave_size;
__syncthreads();
smem[threadIdx.x] = local;
__syncthreads();
// the data within single wave is the same
// but for simplicity, we still use data from each lane.
T v_local = smem[lane_id];
#pragma unroll
for(int i_stage = 1; i_stage < waves; i_stage++)
{
T v_remote = smem[i_stage * wave_size + lane_id];
v_local = reduce_f(v_local, v_remote);
}
return v_local;
}
// kernel entry point
__device__ void operator()(naive_attention_fwd_args args)
{
constexpr int wg_size = get_block_size();
__shared__ char smem[wg_size * 4 * sizeof(float)]; // should enough
int i_dv = blockIdx.x * wg_size + threadIdx.x; // index of hdim_v
int i_sq = blockIdx.y; // index of seqlen_q
int i_batch = blockIdx.z; // index of batch_q * nhead_q
int i_bq = i_batch / args.nhead_q; // index of batch_q
int i_hq = i_batch % args.nhead_q; // index of nhead_q
int i_bk = i_bq / args.batch_ratio_kv;
int i_hk = i_hq / args.nhead_ratio_kv;
void* page_table_ptr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return reinterpret_cast<int*>(args.page_table_ptr) + i_bq * args.max_pages_per_seq;
}
else
{
return nullptr;
}
}();
auto q_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<QType, QLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim, args.q_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return addresser<QType, QLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim, args.q_ptr};
}
}();
auto k_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<KType, KLayout>{
args.batch_kv, args.seqlen_kv, args.nhead_kv, args.hdim, args.k_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return page_addresser<KType, KLayout>{
args.page_size, args.nhead_kv, args.hdim, args.k_ptr, page_table_ptr};
}
}();
auto v_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<VType, VLayout>{
args.batch_kv, args.seqlen_kv, args.nhead_kv, args.hdim_v, args.v_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return page_addresser<VType, VLayout>{
args.page_size, args.nhead_kv, args.hdim_v, args.v_ptr, page_table_ptr};
}
}();
auto o_addr = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return addresser<OType, OLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim_v, args.o_ptr};
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return addresser<OType, OLayout>{
args.batch_q, args.seqlen_q, args.nhead_q, args.hdim_v, args.o_ptr};
}
}();
q_addr.init(i_bq, i_hq);
k_addr.init(i_bk, i_hk);
v_addr.init(i_bk, i_hk);
o_addr.init(i_bq, i_hq);
auto f_max = [](auto x_, auto y_) { return max(x_, y_); };
auto f_sum = [](auto x_, auto y_) { return x_ + y_; };
auto f_absmax_f32 = [](float v_0_, float v_1_) {
float rtn;
asm volatile("v_max_f32 %0, abs(%1), abs(%2)" : "=v"(rtn) : "v"(v_0_), "v"(v_1_));
return rtn;
};
int seqlen_kv = [&]() {
if constexpr(Traits::variation == naive_attention_variation_enum::FLASH_BATCHED)
{
return args.seqlen_kv;
}
else if constexpr(Traits::variation == naive_attention_variation_enum::DECODE_PAGED)
{
return reinterpret_cast<int*>(args.context_len_ptr)[i_bq];
}
}();
SoftmaxType row_max = -numeric<SoftmaxType>::infinity();
SoftmaxType l{0};
AccType o_acc = {0};
int sk_loops = (seqlen_kv + wg_size - 1) / wg_size;
float qf_scale = .0f;
kvscale_addresser<KVScaleType> kvscale_addr{args.nhead_kv, args.hdim, args.kvscale_ptr};
if constexpr(is_kvcache_i8_forward_quant)
{
// AccType is i32 now, seqlen_q = 1, hdim up to 256
float q = 0;
float k_s = 0;
if(static_cast<int>(threadIdx.x) < args.hdim)
{
q = type_convert<float>(q_addr.load(0, threadIdx.x));
k_s = type_convert<float>(kvscale_addr.load(i_hk, threadIdx.x, 0));
}
// 1) we apply the k scale to q
float q_forwarded = q * k_s;
// 2) apply smooth-quant
// find absmax
float qf_max = wave_reduce(q_forwarded, f_absmax_f32);
qf_max = cross_wave_reduce(qf_max, f_absmax_f32, reinterpret_cast<float*>(smem));
// per-token scale
qf_scale = qf_max / 127.0;
// devide by scale
q = q / qf_scale;
// fp32->i8
int8_t quantized_q = static_cast<int8_t>(q);
__syncthreads();
reinterpret_cast<int8_t*>(smem)[threadIdx.x] = quantized_q;
__syncthreads();
// after above process, we have 2 data
// 1) int8 q data stored in smem(no need to reload)
// 2) per-token scale qf_scale, to be mul after 1st gemm
}
for(int i_loop1 = 0; i_loop1 < sk_loops; i_loop1++)
{
int i_sk = i_loop1 * wg_size + threadIdx.x;
// gemm-1
SoftmaxType s_softmax = -numeric<SoftmaxType>::infinity();
if(i_sk < seqlen_kv)
{
AccType s_acc{0}; // clear for every loop
for(auto i_dq = 0; i_dq < args.hdim; i_dq++)
{
if constexpr(is_kvcache_i8_forward_quant)
{
int8_t q = reinterpret_cast<int8_t*>(smem)[i_dq];
auto k = k_addr.load(i_sk, i_dq);
s_acc += type_convert<AccType>(q) * type_convert<AccType>(k);
}
else
{
auto q = q_addr.load(i_sq, i_dq); // q will have duplicate load
auto k = k_addr.load(i_sk, i_dq);
s_acc += type_convert<AccType>(q) * type_convert<AccType>(k);
}
}
// scale
s_softmax = type_convert<SoftmaxType>(s_acc);
s_softmax *=
type_convert<SoftmaxType>(args.scale_s * ck_tile::log2e_v<SoftmaxType>);
if constexpr(is_kvcache_i8_forward_quant)
{
s_softmax *= qf_scale; // post scale the per-token factor
}
}
// s->p
float pf_scale = 0.; // used for i8 quant
{
// softmax, find max
SoftmaxType old_max = row_max;
SoftmaxType cur_max = wave_reduce(s_softmax, f_max);
cur_max = cross_wave_reduce(cur_max, f_max, reinterpret_cast<SoftmaxType*>(smem));
row_max = max(old_max, cur_max); // update row_max
// softmax, exp(i_elem - max)
SoftmaxType p_compute = __builtin_amdgcn_exp2f(s_softmax - row_max);
// compute exp_sum
SoftmaxType row_sum = wave_reduce(p_compute, f_sum);
row_sum = cross_wave_reduce(row_sum, f_sum, reinterpret_cast<SoftmaxType*>(smem));
// l, pre-scall o_acc
SoftmaxType tmp = __builtin_amdgcn_exp2f(old_max - row_max);
l = tmp * l + row_sum;
o_acc = type_convert<AccType>(type_convert<SoftmaxType>(o_acc) * tmp);
// prepare the p_compute into smem, to let every thread read same p_compute and do
// 2nd gemm
if constexpr(is_kvcache_i8_forward_quant)
{
float v_s = 0;
if(static_cast<int>(threadIdx.x) < args.hdim_v)
{
v_s = type_convert<float>(kvscale_addr.load(i_hk, threadIdx.x, 1));
}
// 1) we apply the v scale to p
float p_forwarded = p_compute * v_s;
// 2) apply smooth-quant
// find absmax
float pf_max = wave_reduce(p_forwarded, f_absmax_f32);
pf_max =
cross_wave_reduce(pf_max, f_absmax_f32, reinterpret_cast<float*>(smem));
// per-token scale
pf_scale = pf_max / 127.0;
// devide by scale
p_compute = p_compute / pf_scale;
// fp32->i8
int8_t quantized_p = static_cast<int8_t>(p_compute);
__syncthreads();
reinterpret_cast<int8_t*>(smem)[threadIdx.x] = quantized_p;
__syncthreads();
// after above process, we have 2 data
// 1) int8 p data stored in smem(no need to reload)
// 2) per-token scale pf_scale, to be mul after 2nd gemm
}
else
{
__syncthreads();
reinterpret_cast<PType*>(smem)[threadIdx.x] = type_convert<PType>(p_compute);
__syncthreads();
}
}
// gemm-2, simple loop over vector by vector
constexpr int gemm_2_loop = wg_size / p_vec_elem;
{
AccType o_acc_local = {0};
int sk_start = i_loop1 * wg_size; // we start from the first seqlen_kv element
for(int i_loop2 = 0; i_loop2 < gemm_2_loop; i_loop2++)
{
p_vec_type p_vec = reinterpret_cast<p_vec_type*>(smem)[i_loop2];
#pragma unroll
for(int i_j = 0; i_j < p_vec_elem; i_j++)
{
int sv_offset = i_loop2 * p_vec_elem + i_j;
int i_sv = sk_start + sv_offset;
VType v = 0.f;
if(i_dv < args.hdim_v && i_sv < seqlen_kv)
{
v = v_addr.load(i_sv, i_dv);
}
o_acc_local += type_convert<AccType>(p_vec[i_j]) * type_convert<AccType>(v);
}
}
if constexpr(is_kvcache_i8_forward_quant)
{
// apply pr scale to local acc
o_acc_local =
type_convert<AccType>(type_convert<float>(o_acc_local) * pf_scale);
}
o_acc += o_acc_local;
}
}
// post scale o_acc
{
SoftmaxType tmp = l == 0.f ? 0.f : 1.f / l; // in case masking
o_acc = type_convert<AccType>(type_convert<SoftmaxType>(o_acc) * tmp);
}
// store O
if(i_dv < args.hdim_v)
o_addr.store(type_convert<OType>(o_acc), i_sq, i_dv);
}
};
#define CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_() \
{ \
using ktraits_ = \
naive_attention_fwd_kernel_traits<static_cast<naive_attention_variation_enum>( \
variation_)>; \
using k_ = naive_attention_fwd_kernel<q_type_, \
k_type_, \
v_type_, \
o_type_, \
acc_type_, \
q_layout_, \
k_layout_, \
v_layout_, \
o_layout_, \
ktraits_>; \
dim3 grids = k_::get_grid_size(a); \
r = ck_tile::launch_kernel(s, \
ck_tile::make_kernel(k_{}, grids, k_::get_block_size(), 0, a)); \
}
#define CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_() \
if(t.variation == 0 && t.q_layout == "bshd" && t.k_layout == "bshd" && t.v_layout == "bshd" && \
t.o_layout == "bshd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto k_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto v_layout_ = naive_attention_layout_enum::BSHD; \
constexpr auto o_layout_ = naive_attention_layout_enum::BSHD; \
constexpr int variation_ = 0; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
} \
else if(t.variation == 0 && t.q_layout == "bhsd" && t.k_layout == "bhsd" && \
t.v_layout == "bhsd" && t.o_layout == "bhsd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto k_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto v_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto o_layout_ = naive_attention_layout_enum::BHSD; \
constexpr int variation_ = 0; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
} \
else if(t.variation == 2 && t.q_layout == "bhsd" && t.k_layout == "phdsx" && \
t.v_layout == "phds" && t.o_layout == "bhsd") \
{ \
constexpr auto q_layout_ = naive_attention_layout_enum::BHSD; \
constexpr auto k_layout_ = naive_attention_layout_enum::PHDSX; \
constexpr auto v_layout_ = naive_attention_layout_enum::PHDS; \
constexpr auto o_layout_ = naive_attention_layout_enum::BHSD; \
constexpr int variation_ = 2; \
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_(); \
}
//
CK_TILE_HOST float naive_attention_fwd(naive_attention_fwd_traits t,
naive_attention_fwd_args a,
ck_tile::stream_config s)
{
float r = -1;
// TODO: do not explicitly create too much instance!
if(t.q_type == "fp16" && t.k_type == "fp16" && t.v_type == "fp16" && t.o_type == "fp16")
{
using q_type_ = fp16_t;
using k_type_ = fp16_t;
using v_type_ = fp16_t;
using o_type_ = fp16_t;
using acc_type_ = float;
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "bf16" && t.k_type == "bf16" && t.v_type == "bf16" && t.o_type == "bf16")
{
using q_type_ = bf16_t;
using k_type_ = bf16_t;
using v_type_ = bf16_t;
using o_type_ = bf16_t;
using acc_type_ = float;
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "bf16" && t.k_type == "int8" && t.v_type == "int8" && t.o_type == "bf16")
{
using q_type_ = bf16_t;
using k_type_ = int8_t;
using v_type_ = int8_t;
using o_type_ = bf16_t;
using acc_type_ = int32_t; // NOTE!
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
else if(t.q_type == "fp16" && t.k_type == "int8" && t.v_type == "int8" && t.o_type == "fp16")
{
using q_type_ = fp16_t;
using k_type_ = int8_t;
using v_type_ = int8_t;
using o_type_ = fp16_t;
using acc_type_ = int32_t; // NOTE!
CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_();
}
return r;
}
#undef CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_LAOYUT_
#undef CK_TILE_DISPATCH_NAIVE_ATTEN_FWD_INTERNAL_
} // namespace ck_tile
include/ck_tile/remod.py
View file @ 75535dd8
...@@ -7,6 +7,7 @@ import copy ...@@ -7,6 +7,7 @@ import copy
NS = 'ck_tile' NS = 'ck_tile'
OPS = 'ops' OPS = 'ops'
REF = 'ref'
OPS_COMMON = 'common' # common header will be duplicated into ops/* other module OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
HEADER_COMMON = f"""// SPDX-License-Identifier: MIT HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
...@@ -29,6 +30,9 @@ class submodule_t: ...@@ -29,6 +30,9 @@ class submodule_t:
def push(self, f): def push(self, f):
if len(f.parents) != 1: # ignore ./xxx.hpp if len(f.parents) != 1: # ignore ./xxx.hpp
mod = get_module(f) mod = get_module(f)
# ref is supposed to include one header on demand
if mod == REF:
return
if mod == OPS: if mod == OPS:
if mod not in self.m.keys(): if mod not in self.m.keys():
self.m[mod] = dict() self.m[mod] = dict()
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment