Implement EmbeddingBag

[cognaiger9]

• Add EmbeddingBag operation with forward kernels.
• Add driver and gtest for kernels.
• MIOpen performs better if:
  • Mode: Max
  • Mode: Mean or Sum, when the tensor type is float and all tensors are contiguous and number of elements in the output exceeds 2^19
• Note:
  • Forward solver only works with 2D tensor

Average improvement over ROCm

type	fwd
float16	1.38
float	1.3
bfloat16	1.4

Detail Benchmark

float16 (mode max)

op_name	dtype	input size	weight size	cont	direction	ROCm	MIOpen	Improvement
EmbeddingBag	float16	[32 655]	[100 256]	cont	fwd	320271	253759	1.26
EmbeddingBag	float16	[512 512]	[256 1024]	cont	fwd	1198653	846078	1.42
EmbeddingBag	float16	[512 512]	[256 1024]	noncont	fwd	1210990	1012089	1.20
EmbeddingBag	float16	[512 512]	[512 2048]	cont	fwd	2385852	1520730	1.57
EmbeddingBag	float16	[512 512]	[512 2048]	noncont	fwd	2394267	1838890	1.30
EmbeddingBag	float16	[1024 512]	[256 1024]	cont	fwd	2391292	1555550	1.54
EmbeddingBag	float16	[1024 512]	[512 2048]	cont	fwd	4765832	2955600	1.61
EmbeddingBag	float16	[1024 512]	[512 2048]	noncont	fwd	4791704	3989040	1.20
EmbeddingBag	float16	[768 768]	[256 1024]	cont	fwd	2684508	1778770	1.51
EmbeddingBag	float16	[768 768]	[256 1024]	noncont	fwd	2706427	2111200	1.28
EmbeddingBag	float16	[768 768]	[512 2048]	cont	fwd	5352231	3353980	1.60
EmbeddingBag	float16	[768 768]	[512 2048]	noncont	fwd	5376823	4098639	1.31
EmbeddingBag	float16	[256 256]	[512 2048]	cont	fwd	607119	422861	1.44
EmbeddingBag	float16	[256 256]	[512 2048]	noncont	fwd	612143	508318	1.20
EmbeddingBag	float16	[16 255]	[100 256]	cont	fwd	129472	108169	1.20

float32 (mode max)

op_name	dtype	input size	weight size	cont	direction	ROCm	MIOpen	Improvement
EmbeddingBag	float32	[512 512]	[256 1024]	cont	fwd	1191598	863910	1.38
EmbeddingBag	float32	[512 512]	[256 1024]	noncont	fwd	1191582	1029700	1.16
EmbeddingBag	float32	[512 512]	[512 2048]	cont	fwd	2364124	1556490	1.52
EmbeddingBag	float32	[512 512]	[512 2048]	noncont	fwd	2376844	1863040	1.28
EmbeddingBag	float32	[1024 512]	[256 1024]	cont	fwd	2381244	1587020	1.50
EmbeddingBag	float32	[1024 512]	[512 2048]	cont	fwd	4769672	3037880	1.57
EmbeddingBag	float32	[1024 512]	[512 2048]	noncont	fwd	4799640	3896120	1.23
EmbeddingBag	float32	[768 768]	[256 1024]	cont	fwd	2664284	1819250	1.46
EmbeddingBag	float32	[768 768]	[256 1024]	noncont	fwd	2684236	2146680	1.25
EmbeddingBag	float32	[768 768]	[512 2048]	cont	fwd	5335863	3436330	1.55
EmbeddingBag	float32	[768 768]	[512 2048]	noncont	fwd	5371095	4152190	1.29
EmbeddingBag	float32	[256 256]	[512 2048]	cont	fwd	602367	435573	1.38
EmbeddingBag	float32	[256 256]	[512 2048]	noncont	fwd	607999	517136	1.18
EmbeddingBag	float32	[128 512]	[512 2048]	cont	fwd	600879	528356	1.14

bfloat16 (mode max)

op_name	dtype	input size	weight size	cont	direction	ROCm	MIOpen	Improvement
EmbeddingBag	bfloat16	[512 512]	[256 1024]	cont	fwd	1188750	834043	1.43
EmbeddingBag	bfloat16	[512 512]	[256 1024]	noncont	fwd	1188286	1010610	1.18
EmbeddingBag	bfloat16	[512 512]	[512 2048]	cont	fwd	2360604	1491550	1.58
EmbeddingBag	bfloat16	[512 512]	[512 2048]	noncont	fwd	2360636	1816100	1.30
EmbeddingBag	bfloat16	[1024 512]	[256 1024]	cont	fwd	2373468	1527450	1.55
EmbeddingBag	bfloat16	[1024 512]	[512 2048]	cont	fwd	4721576	2911920	1.62
EmbeddingBag	bfloat16	[1024 512]	[512 2048]	noncont	fwd	4735288	4048420	1.17
EmbeddingBag	bfloat16	[768 768]	[100 256]	cont	fwd	637519	525137	1.21
EmbeddingBag	bfloat16	[768 768]	[256 1024]	cont	fwd	2662716	1746410	1.52
EmbeddingBag	bfloat16	[768 768]	[256 1024]	noncont	fwd	2681467	2100390	1.28
EmbeddingBag	bfloat16	[768 768]	[512 2048]	cont	fwd	5294503	3315640	1.60
EmbeddingBag	bfloat16	[768 768]	[512 2048]	noncont	fwd	5323143	4110299	1.30
EmbeddingBag	bfloat16	[256 256]	[512 2048]	cont	fwd	600703	416995	1.44

float32 (mode mean)

op_name	dtype	size	direction	cont	direction	ROCm	MIOpen	Improvement
EmbeddingBag	float32	[512 512]	[256 1024]	cont	fwd	1154688	955314	1.21
EmbeddingBag	float32	[512 512]	[512 2048]	cont	fwd	2307201	1764120	1.31
EmbeddingBag	float32	[1024 512]	[256 1024]	cont	fwd	2303857	1798470	1.28
EmbeddingBag	float32	[1024 512]	[512 2048]	cont	fwd	4597474	3431640	1.34
EmbeddingBag	float32	[768 768]	[512 2048]	cont	fwd	2588381	2049459	1.26
EmbeddingBag	float32	[768 768]	[512 2048]	cont	fwd	5169386	3869800	1.34
EmbeddingBag	float32	[256 256]	[512 2048]	cont	fwd	589464	481217	1.22
EmbeddingBag	float32	[16 255]	[512 2048]	cont	fwd	125583	114259	1.10
EmbeddingBag	float32	[16 255]	[512 2048]	cont	fwd	142318	126775	1.12
EmbeddingBag	float32	[128 128]	[100 256]	cont	fwd	72623	65636	1.11

[BradPepersAMD]

MIOpen is moving to the new monorepo setup and all older unmerged PR's are being closed. Please re-open this as part of the new repo if these changes are still needed.