[mlir-gen] Add Float16 type support (#901)

Allows IR generation with f16 data type. The new type is mostly relevant for GPU testing as f16 is a common target for hardware accelerated operations. Required changes are propagated around as extra differentiation is now needed to select between f16 and bf16.
plaidml · Mar 12, 2024 · b04f662 · b04f662
1 parent 887d4a2
commit b04f662
Show file tree

Hide file tree

Showing 52 changed files with 327 additions and 310 deletions.
diff --git a/benchmarks/config/base/base.json b/benchmarks/config/base/base.json
@@ -31,42 +31,42 @@
     },
     "gemm_fp32_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=32 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=f32 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": []
     },
     "gemm_bf16_dp2_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=2" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=2" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "avx2" ]
     },
     "gemm_bf16_dp4_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=4" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=4" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "svebf16" ]
     },
     "mlp_fp32_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-width=32 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-type=f32 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": []
     },
     "mlp_bf16_dp2_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-width=16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=2" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=2" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "avx2" ]
     },
     "mlp_bf16_dp4_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-width=16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=4" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024 --tiles=32,32,32 --vnni=4" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "svebf16" ]
@@ -76,28 +76,28 @@
   "gemm_models": {
     "fp32_3x1024_const_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=32 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=f32 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fp32_3x1024_args_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --float-width=32 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --float-type=f32 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "bf16_3x1024_const_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=16 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100"],
       "extensions": [ "(avx2|asimd)" ]
     },
     "bf16_3x1024_args_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --float-width=16 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100"],
       "extensions": [ "(avx2|asimd)" ]
@@ -107,28 +107,28 @@
   "mlp_models": {
     "fp32_3x1024_const_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --float-width=32 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --float-type=f32 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fp32_3x1024_args_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "bf16_3x1024_const_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-width=16 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=const --bias --relu --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100"],
       "extensions": [ "(avx2|asimd)" ]
     },
     "bf16_3x1024_args_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --batch=256 --layers=1024,1024,1024,1024" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --batch=256 --layers=1024,1024,1024,1024" ],
       "environment": {},
       "flags": [ "-n", "100"],
       "extensions": [ "(avx2|asimd)" ]

diff --git a/benchmarks/config/fc/1024x1024x512.json b/benchmarks/config/fc/1024x1024x512.json
@@ -37,14 +37,14 @@
   "fc_1024x1024x512_fp32_mlir": {
     "fc_fp32_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "1" },
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_fp32_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -54,14 +54,14 @@
   "fc_1024x1024x512_bf16_dp2_mlir": {
     "fc_bf16_dp2_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_bf16_dp2_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -71,14 +71,14 @@
   "fc_1024x1024x512_bf16_dp4_mlir": {
     "fc_bf16_dp4_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(svebf16)" ]
     },
     "fc_bf16_dp4_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=512,1024 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(svebf16)" ]

diff --git a/benchmarks/config/fc/1024x2560x1024.json b/benchmarks/config/fc/1024x2560x1024.json
@@ -37,14 +37,14 @@
   "fc_1024x2560x1024_fp32_mlir": {
     "fc_fp32_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "1" },
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_fp32_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -54,14 +54,14 @@
   "fc_1024x2560x1024_bf16_dp2_mlir": {
     "fc_bf16_dp2_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_bf16_dp2_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -71,14 +71,14 @@
   "fc_1024x2560x1024_bf16_dp4_mlir": {
     "fc_bf16_dp4_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(svebf16)" ]
     },
     "fc_bf16_dp4_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=1024,2560 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(svebf16)" ]

diff --git a/benchmarks/config/fc/1024x352x512.json b/benchmarks/config/fc/1024x352x512.json
@@ -37,14 +37,14 @@
   "fc_1024x352x512_fp32_mlir": {
     "fc_fp32_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": { "OMP_NUM_THREADS": "1" },
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_fp32_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -54,14 +54,14 @@
   "fc_1024x352x512_bf16_dp2_mlir": {
     "fc_bf16_dp2_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_bf16_dp2_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -71,14 +71,14 @@
   "fc_1024x352x512_bf16_dp4_mlir": {
     "fc_bf16_dp4_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(svebf16)" ]
     },
     "fc_bf16_dp4_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=512,352 --tiles=32,32,32" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(svebf16)" ]

diff --git a/benchmarks/config/fc/1024x512x256.json b/benchmarks/config/fc/1024x512x256.json
@@ -37,14 +37,14 @@
   "fc_1024x512x256_fp32_mlir": {
     "fc_fp32_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "1" },
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_fp32_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=32 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=f32 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -54,14 +54,14 @@
   "fc_1024x512x256_bf16_dp2_mlir": {
     "fc_bf16_dp2_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(avx2|asimd)" ]
     },
     "fc_bf16_dp2_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=2 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=2 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(avx2|asimd)" ]
@@ -71,14 +71,14 @@
   "fc_1024x512x256_bf16_dp4_mlir": {
     "fc_bf16_dp4_single_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": {},
       "flags": [ "-n", "100" ],
       "extensions": [ "(svebf16)" ]
     },
     "fc_bf16_dp4_omp_16_mlir": {
       "type": "IR-GEN",
-      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-width=16 --vnni=4 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
+      "benchmark": [ "mlir-gen", "--kernel=args --bias --relu --float-type=bf16 --vnni=4 --batch=1024 --layers=256,512 --tiles=64,64,64" ],
       "environment": { "OMP_NUM_THREADS": "16", "KMP_AFFINITY": "granularity=fine,verbose,compact,1,0" },
       "flags": [ "-n", "100", "-run-args='-def-parallel'" ],
       "extensions": [ "(svebf16)" ]