tjh.dev
nixpkgs mirror (for testing)
github.com/NixOS/nixpkgs
nix
1{
2 lib,
3 clr,
4 composable_kernel_base,
5}:
6
7let
8 inherit (composable_kernel_base) miOpenReqLibsOnly;
9 parts = {
10 _mha = {
11 enabled = !miOpenReqLibsOnly;
12 # mha takes ~3hrs on 64 cores on an EPYC milan system at ~2.5GHz
13 # big-parallel builders are one gen newer and clocked ~30% higher but only 24 cores
14 # Should be <10h timeout but might be cutting it close
15 # TODO: work out how to split this into smaller chunks instead of all 3k mha instances together
16 # mha_0,1,2, search ninja target file for the individual instances, split by the index?
17 # TODO: can we prune the generated instances down to only what in practice are used with popular models
18 # when using flash-attention + MHA kernels?
19 targets = [
20 "device_mha_instance"
21 ];
22 extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
23 };
24 gemm_multiply_multiply = {
25 enabled = !miOpenReqLibsOnly;
26 targets = [
27 "device_gemm_multiply_multiply_instance"
28 ];
29 extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
30 onlyFor = [
31 "gfx942"
32 "gfx950"
33 ];
34 };
35 gemm_multiply_multiply_wp = {
36 enabled = !miOpenReqLibsOnly;
37 targets = [
38 "device_gemm_multiply_multiply_wp_instance"
39 ];
40 extraCmakeFlags = [ "-DHIP_CLANG_NUM_PARALLEL_JOBS=2" ];
41 onlyFor = [
42 "gfx942"
43 "gfx950"
44 ];
45 };
46 grouped_conv_bwd = {
47 targets = [
48 "device_grouped_conv1d_bwd_weight_instance"
49 "device_grouped_conv2d_bwd_data_instance"
50 "device_grouped_conv2d_bwd_weight_instance"
51 ];
52 };
53 grouped_conv_fwd = {
54 targets = [
55 "device_grouped_conv1d_fwd_instance"
56 "device_grouped_conv2d_fwd_instance"
57 "device_grouped_conv2d_fwd_bias_clamp_instance"
58 "device_grouped_conv2d_fwd_clamp_instance"
59 "device_grouped_conv2d_fwd_dynamic_op_instance"
60 ];
61 };
62 grouped_conv_bwd_3d1 = {
63 targets = [
64 "device_grouped_conv3d_bwd_data_instance"
65 "device_grouped_conv3d_bwd_data_bilinear_instance"
66 "device_grouped_conv3d_bwd_data_scale_instance"
67 ];
68 };
69 grouped_conv_bwd_3d2 = {
70 targets = [
71 "device_grouped_conv3d_bwd_weight_instance"
72 "device_grouped_conv3d_bwd_weight_bilinear_instance"
73 "device_grouped_conv3d_bwd_weight_scale_instance"
74 ];
75 };
76 grouped_conv_fwd_3d1 = {
77 targets = [
78 "device_grouped_conv3d_fwd_instance"
79 "device_grouped_conv3d_fwd_clamp_instance"
80 "device_grouped_conv3d_fwd_bias_clamp_instance"
81 "device_grouped_conv3d_fwd_bilinear_instance"
82 "device_grouped_conv3d_fwd_convinvscale_instance"
83 "device_grouped_conv3d_fwd_convscale_instance"
84 "device_grouped_conv3d_fwd_convscale_add_instance"
85 ];
86 };
87 grouped_conv_fwd_3d2 = {
88 targets = [
89 "device_grouped_conv3d_fwd_convscale_relu_instance"
90 "device_grouped_conv3d_fwd_dynamic_op_instance"
91 "device_grouped_conv3d_fwd_scale_instance"
92 "device_grouped_conv3d_fwd_scaleadd_ab_instance"
93 "device_grouped_conv3d_fwd_scaleadd_scaleadd_relu_instance"
94 ];
95 };
96 grouped_conv_fwd_nd = {
97 targets = [
98 "device_grouped_convnd_bwd_weight_instance"
99 ];
100 };
101 batched_gemm1 = {
102 enabled = !miOpenReqLibsOnly;
103 targets = [
104 "device_batched_gemm_instance"
105 "device_batched_gemm_b_scale_instance"
106 "device_batched_gemm_multi_d_instance"
107 "device_batched_gemm_add_relu_gemm_add_instance"
108 "device_batched_gemm_bias_permute_instance"
109 "device_batched_gemm_gemm_instance"
110 "device_batched_gemm_reduce_instance"
111 "device_batched_gemm_softmax_gemm_instance"
112 ];
113 };
114 batched_gemm2 = {
115 enabled = !miOpenReqLibsOnly;
116 targets = [
117 "device_batched_gemm_softmax_gemm_permute_instance"
118 "device_grouped_gemm_instance"
119 "device_grouped_gemm_bias_instance"
120 "device_grouped_gemm_fastgelu_instance"
121 "device_grouped_gemm_fixed_nk_instance"
122 "device_grouped_gemm_fixed_nk_multi_abd_instance"
123 "device_grouped_gemm_tile_loop_instance"
124 ];
125 };
126 gemm_universal1 = {
127 enabled = !miOpenReqLibsOnly;
128 targets = [
129 "device_gemm_universal_instance"
130 "device_gemm_universal_batched_instance"
131 ];
132 };
133 gemm_universal2 = {
134 enabled = !miOpenReqLibsOnly;
135 targets = [
136 "device_gemm_universal_reduce_instance"
137 "device_gemm_universal_streamk_instance"
138 ];
139 };
140 gemm_other1 = {
141 enabled = !miOpenReqLibsOnly;
142 targets = [
143 "device_gemm_instance"
144 "device_gemm_b_scale_instance"
145 "device_gemm_ab_scale_instance"
146 "device_gemm_add_instance"
147 "device_gemm_add_add_fastgelu_instance"
148 "device_gemm_add_fastgelu_instance"
149 "device_gemm_add_multiply_instance"
150 "device_gemm_add_relu_instance"
151 ];
152 };
153 gemm_other2 = {
154 enabled = !miOpenReqLibsOnly;
155 targets = [
156 "device_gemm_add_relu_add_layernorm_instance"
157 "device_gemm_add_silu_instance"
158 "device_gemm_bias_add_reduce_instance"
159 "device_gemm_bilinear_instance"
160 "device_gemm_fastgelu_instance"
161 "device_gemm_multi_abd_instance"
162 "device_gemm_multiply_add_instance"
163 "device_gemm_reduce_instance"
164 "device_gemm_splitk_instance"
165 "device_gemm_streamk_instance"
166 ];
167 };
168 conv = {
169 targets = [
170 "device_conv1d_bwd_data_instance"
171 "device_conv2d_bwd_data_instance"
172 "device_conv2d_fwd_instance"
173 "device_conv2d_fwd_bias_relu_instance"
174 "device_conv2d_fwd_bias_relu_add_instance"
175 "device_conv3d_bwd_data_instance"
176 ];
177 };
178 pool = {
179 enabled = !miOpenReqLibsOnly;
180 targets = [
181 "device_avg_pool2d_bwd_instance"
182 "device_avg_pool3d_bwd_instance"
183 "device_pool2d_fwd_instance"
184 "device_pool3d_fwd_instance"
185 "device_max_pool_bwd_instance"
186 ];
187 };
188 other0 = {
189 targets = [
190 "device_quantization_instance"
191 ];
192 };
193 other1 = {
194 enabled = !miOpenReqLibsOnly;
195 targets = [
196 "device_batchnorm_instance"
197 "device_contraction_bilinear_instance"
198 "device_contraction_scale_instance"
199 "device_elementwise_instance"
200 "device_elementwise_normalization_instance"
201 ];
202 };
203 other2 = {
204 enabled = !miOpenReqLibsOnly;
205 targets = [
206 "device_column_to_image_instance"
207 "device_image_to_column_instance"
208 "device_permute_scale_instance"
209 "device_reduce_instance"
210 ];
211 };
212 other3 = {
213 enabled = !miOpenReqLibsOnly;
214 targets = [
215 "device_normalization_bwd_data_instance"
216 "device_normalization_bwd_gamma_beta_instance"
217 "device_normalization_fwd_instance"
218 "device_softmax_instance"
219 "device_transpose_instance"
220 ];
221 };
222 };
223 tensorOpBuilder =
224 {
225 part,
226 targets,
227 extraCmakeFlags ? [ ],
228 requiredSystemFeatures ? [ "big-parallel" ],
229 enabled ? true,
230 onlyFor ? [ ],
231 }:
232 let
233 supported =
234 enabled
235 && (onlyFor == [ ] || (lib.lists.intersectLists composable_kernel_base.gpuTargets onlyFor) != [ ]);
236 in
237 if supported then
238 (composable_kernel_base.overrideAttrs (old: {
239 inherit requiredSystemFeatures;
240 pname = "composable_kernel${clr.gpuArchSuffix}-${part}";
241 makeTargets = targets;
242 preBuild = ''
243 echo "Building ${part}"
244 makeFlagsArray+=($makeTargets)
245 substituteInPlace $(find ./ -name "Makefile" -type f) \
246 --replace-fail '.NOTPARALLEL:' '.UNUSED_NOTPARALLEL:'
247 '';
248
249 # Compile parallelism adjusted based on available RAM
250 # Never uses less than NIX_BUILD_CORES/4, never uses more than NIX_BUILD_CORES
251 # CK uses an unusually high amount of memory per core in the build step
252 # Nix/nixpkgs doesn't really have any infra to tell it that this build is unusually memory hungry
253 # So, bodge. Otherwise you end up having to build all of ROCm with a low core limit when
254 # it's only this package that has trouble.
255 preConfigure = old.preConfigure or "" + ''
256 MEM_GB_TOTAL=$(awk '/MemTotal/ { printf "%d \n", $2/1024/1024 }' /proc/meminfo)
257 MEM_GB_AVAILABLE=$(awk '/MemAvailable/ { printf "%d \n", $2/1024/1024 }' /proc/meminfo)
258 APPX_GB=$((MEM_GB_AVAILABLE > MEM_GB_TOTAL ? MEM_GB_TOTAL : MEM_GB_AVAILABLE))
259 MAX_CORES=$((1 + APPX_GB/3))
260 MAX_CORES=$((MAX_CORES < NIX_BUILD_CORES/3 ? NIX_BUILD_CORES/3 : MAX_CORES))
261 export NIX_BUILD_CORES="$((NIX_BUILD_CORES > MAX_CORES ? MAX_CORES : NIX_BUILD_CORES))"
262 echo "Picked new core limit NIX_BUILD_CORES=$NIX_BUILD_CORES based on available mem: $APPX_GB GB"
263 cmakeFlagsArray+=(
264 "-DCK_PARALLEL_COMPILE_JOBS=$NIX_BUILD_CORES"
265 )
266 '';
267 cmakeFlags = old.cmakeFlags ++ extraCmakeFlags;
268 # Early exit after build phase with success, skips fixups etc
269 # Will get copied back into /build of the final CK
270 postBuild = ''
271 find . -name "*.o" -type f | while read -r file; do
272 mkdir -p "$out/$(dirname "$file")"
273 cp --reflink=auto "$file" "$out/$file"
274 done
275 exit 0
276 '';
277 meta = old.meta // {
278 broken = false;
279 };
280 }))
281 else
282 null;
283 composable_kernel_parts = builtins.mapAttrs (
284 part: targets: tensorOpBuilder (targets // { inherit part; })
285 ) parts;
286in
287
288composable_kernel_base.overrideAttrs (
289 finalAttrs: old: {
290 pname = "composable_kernel${clr.gpuArchSuffix}";
291 parts_dirs = builtins.filter (x: x != null) (builtins.attrValues composable_kernel_parts);
292 disallowedReferences = builtins.filter (x: x != null) (builtins.attrValues composable_kernel_parts);
293 preBuild = ''
294 for dir in $parts_dirs; do
295 find "$dir" -type f -name "*.o" | while read -r file; do
296 # Extract the relative path by removing the output directory prefix
297 rel_path="''${file#"$dir/"}"
298
299 # Create parent directory if it doesn't exist
300 mkdir -p "$(dirname "$rel_path")"
301
302 # Copy the file back to its original location, give it a future timestamp
303 # so make treats it as up to date
304 cp --reflink=auto --no-preserve=all "$file" "$rel_path"
305 touch -d "now +10 hours" "$rel_path"
306 done
307 done
308 '';
309 passthru = old.passthru // {
310 parts = composable_kernel_parts;
311 };
312 meta = old.meta // {
313 # Builds without any gfx9 fail
314 broken = !finalAttrs.passthru.anyGfx9Target;
315 };
316 }
317)