Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Improve heuristic for memcopy tile sizes #172

Merged
merged 1 commit into from
Nov 14, 2023
Merged

Improve heuristic for memcopy tile sizes #172

merged 1 commit into from
Nov 14, 2023

Conversation

thomasfaingnaert
Copy link
Member

Example:

include("configs/configs.jl")

using Test
using CUDA
using GemmKernels

M = N = K = 256
AB_type = Float16
CD_type = Float32
transpose_a = false
transpose_b = false
BLOCK_M, BLOCK_N, BLOCK_K = 64, 128, 16
WARPS_M, WARPS_N = 4, 2
zero_c = false
OP_M, OP_N, OP_K = 16, 16, 16

cf = @get_wmma_config

c_h, a, b, c, d = generate_inputs(cf)

run_gemm(cf, a, b, c, d)

@test verify(cf, c_h, d)

By default, e.g. mem_a_thread is 8 FP16 elements per thread (to get 16 bytes per memory instruction), but this means that each warp handles 32 * 8 = 256 elements per iteration, or 256 * 4 * 2 = 2048 elements for all 8 warps. However, there are only BLOCK_M * BLOCK_K = 1024 elements in a CTA tile, so a_frag_i is 0.

This PR fixes this case by providing a better default, and additionally checking whether there is at least one iteration and throwing a ConfigError if there is not.

Base automatically changed from tf/refactor-configs to master November 14, 2023 15:57
@maleadt
Copy link
Member

maleadt commented Nov 14, 2023

Benchmark results for commit 57ee67e (comparing to 8a73476):

test master PR Δmin
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 7.43 μs ± 2.02% (7.15 … 7.87 μs) 8.47 μs ± 2.0% (8.11 … 8.82 μs) +13.3% ❌
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 124.0 μs ± 0.146% (124.0 … 124.0 μs) 121.0 μs ± 4.75% (109.0 … 124.0 μs) -12.0% ✅
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 222.0 μs ± 4.34% (206.0 … 237.0 μs) 230.0 μs ± 1.06% (224.0 … 237.0 μs) +8.6% ❌
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (128×128) · (128×128) (NN) OP (16, 16, 16) 11.2 μs ± 1.12% (11.0 … 11.7 μs) 9.96 μs ± 1.14% (9.54 … 10.3 μs) -13.0% ✅
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 125.0 μs ± 0.129% (125.0 … 126.0 μs) 124.0 μs ± 3.73% (110.0 … 126.0 μs) -11.9% ✅
WMMA diagonal GEMM diag(Float16)*Float16=Float32 (128×128) · (128×128) (NT) OP (16, 16, 16) 11.2 μs ± 1.28% (11.0 … 11.7 μs) 9.93 μs ± 1.37% (9.54 … 10.3 μs) -13.0% ✅
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 112.0 μs ± 0.153% (112.0 … 113.0 μs) 127.0 μs ± 0.133% (127.0 … 128.0 μs) +13.4% ❌
Tropical GEMM Float32*Float32=Float32 (128×256) · (256×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 124.0 μs ± 3.15% (110.0 … 126.0 μs) 125.0 μs ± 0.119% (124.0 … 126.0 μs) +13.5% ❌
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 8, 1), base shape (4, 8, 1) 15.3 μs ± 1.08% (15.0 … 15.5 μs) 13.4 μs ± 1.14% (13.1 … 13.8 μs) -12.7% ✅
Tropical GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 55.5 μs ± 0.209% (55.1 … 56.0 μs) 63.1 μs ± 0.247% (62.7 … 63.7 μs) +13.9% ❌
WMMA Complex GEMM Float16*Float16=Float32 (128×128) · (128×128) (TT) OP (16, 16, 16) 9.5 μs ± 1.5% (9.06 … 9.78 μs) 8.38 μs ± 1.97% (8.11 … 8.82 μs) -10.5% ✅

Comparison with baseline

test GemmKernels Baseline %
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 15.0 μs ± 1.13% (14.8 … 15.3 μs) 7.21 μs ± 1.72% (6.91 … 7.63 μs) 46.8
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 32, 1), base shape (1, 32, 1) 16.1 μs ± 0.968% (15.7 … 16.2 μs) 4.19 μs ± 3.65% (4.05 … 4.53 μs) 25.8
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 223.0 μs ± 4.03% (207.0 … 235.0 μs) 189.0 μs ± 2.78% (180.0 … 195.0 μs) 86.9
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 7.94 μs ± 1.68% (7.63 … 8.34 μs) 2.14 μs ± 8.84% (1.91 … 2.62 μs) 25.0
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 216.0 μs ± 4.02% (203.0 … 229.0 μs) 186.0 μs ± 2.37% (172.0 … 193.0 μs) 84.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (16, 16, 1), base shape (8, 4, 1) 14.5 μs ± 1.08% (14.3 … 14.8 μs) 4.04 μs ± 3.4% (3.81 … 4.29 μs) 26.7
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 1.85 ms ± 3.45% (1.81 … 2.02 ms) 997.0 μs ± 1.48% (933.0 … 1030.0 μs) 51.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (32, 4, 1), base shape (32, 1, 1) 13.1 μs ± 0.966% (12.6 … 13.4 μs) 4.04 μs ± 3.36% (3.81 … 4.29 μs) 30.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (32, 4, 1), base shape (16, 2, 1) 19.1 μs ± 0.892% (18.8 … 19.3 μs) 4.09 μs ± 2.57% (3.81 … 4.53 μs) 20.3
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 1.84 ms ± 3.11% (1.8 … 2.01 ms) 292.0 μs ± 0.362% (289.0 … 293.0 μs) 16.1
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 46.1 ms ± 0.967% (46.0 … 52.4 ms) 40.8 ms ± 0.0061% (40.8 … 40.8 ms) 88.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (32, 4, 1), base shape (16, 2, 1) 19.1 μs ± 0.927% (18.8 … 19.3 μs) 4.29 μs ± 3.51% (4.05 … 4.53 μs) 21.5
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 288.0 μs ± 3.37% (255.0 … 293.0 μs) 109.0 μs ± 0.156% (109.0 … 110.0 μs) 42.7
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.47 μs ± 2.0% (8.11 … 8.82 μs) 3.78 μs ± 3.35% (3.58 … 4.29 μs) 44.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 15.3 μs ± 1.17% (15.0 … 15.7 μs) 4.09 μs ± 2.59% (3.81 … 4.53 μs) 25.4
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.76 μs ± 1.89% (8.34 … 9.06 μs) 5.19 μs ± 3.16% (5.01 … 5.48 μs) 60.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 8, 1), base shape (4, 8, 1) 15.2 μs ± 1.1% (15.0 … 15.5 μs) 4.3 μs ± 3.39% (4.05 … 4.53 μs) 27.0
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 1.88 ms ± 3.09% (1.85 … 2.05 ms) 1.1 ms ± 1.53% (1.03 … 1.14 ms) 56.1
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 282.0 μs ± 5.52% (256.0 … 292.0 μs) 109.0 μs ± 0.145% (109.0 … 110.0 μs) 42.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 8, 1), base shape (4, 8, 1) 15.2 μs ± 1.09% (15.0 … 15.5 μs) 4.04 μs ± 3.32% (3.81 … 4.29 μs) 25.4
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 364.0 μs ± 5.06% (342.0 … 390.0 μs) 291.0 μs ± 0.159% (290.0 … 294.0 μs) 84.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (16, 16, 1), base shape (4, 8, 1) 15.3 μs ± 0.891% (15.0 … 15.5 μs) 4.09 μs ± 2.49% (3.81 … 4.53 μs) 25.4
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 1.45 ms ± 1.84% (1.43 … 1.53 ms) 1.11 ms ± 1.74% (1.04 … 1.14 ms) 72.5
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.05 μs ± 2.11% (7.87 … 8.34 μs) 2.22 μs ± 7.46% (1.91 … 2.62 μs) 24.2
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 1.87 ms ± 3.31% (1.82 … 2.03 ms) 293.0 μs ± 0.219% (291.0 … 294.0 μs) 16.0
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 130.0 μs ± 0.423% (128.0 … 131.0 μs) 28.7 μs ± 0.442% (28.6 … 29.1 μs) 22.3
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.01 μs ± 1.96% (7.87 … 8.34 μs) 2.18 μs ± 7.89% (1.91 … 2.38 μs) 24.2
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 46.1 ms ± 0.749% (46.0 … 52.4 ms) 40.8 ms ± 0.0059% (40.8 … 40.8 ms) 88.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (32, 4, 1), base shape (32, 1, 1) 12.6 μs ± 1.28% (12.4 … 12.9 μs) 4.19 μs ± 3.7% (4.05 … 4.53 μs) 32.7
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 370.0 μs ± 5.27% (349.0 … 398.0 μs) 291.0 μs ± 0.142% (289.0 … 292.0 μs) 83.0
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 13.0 μs ± 1.14% (12.6 … 13.1 μs) 3.8 μs ± 3.01% (3.58 … 4.29 μs) 28.3
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.42 μs ± 1.9% (8.11 … 8.82 μs) 3.7 μs ± 3.79% (3.58 … 4.05 μs) 44.1
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 370.0 μs ± 4.87% (343.0 … 388.0 μs) 292.0 μs ± 0.291% (289.0 … 293.0 μs) 84.4
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 2.3 ms ± 2.8% (2.26 … 2.56 ms) 291.0 μs ± 0.248% (289.0 … 293.0 μs) 12.8
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 6.91 μs ± 0.782% (6.44 … 7.15 μs) 2.08 μs ± 6.65% (1.91 … 2.38 μs) 29.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 9.04 μs ± 1.86% (8.82 … 9.3 μs) 2.88 μs ± 5.05% (2.62 … 3.1 μs) 29.7
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.3 μs ± 1.15% (12.2 … 12.6 μs) 2.91 μs ± 5.23% (2.62 … 3.1 μs) 21.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 8, 2), base shape (4, 8, 1) 16.0 μs ± 1.05% (15.7 … 16.5 μs) 4.19 μs ± 3.76% (4.05 … 4.53 μs) 25.8
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 16.4 μs ± 1.03% (16.0 … 16.7 μs) 4.2 μs ± 3.75% (4.05 … 4.53 μs) 25.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (32, 4, 1), base shape (32, 1, 1) 31.6 μs ± 0.471% (31.5 … 31.9 μs) 4.08 μs ± 2.56% (3.81 … 4.29 μs) 12.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 13.6 μs ± 0.99% (13.1 … 13.8 μs) 4.04 μs ± 3.46% (3.81 … 4.29 μs) 29.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 32, 1), base shape (1, 32, 1) 15.7 μs ± 1.09% (15.5 … 16.0 μs) 4.09 μs ± 2.49% (3.81 … 4.53 μs) 24.6
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 18.1 μs ± 0.877% (17.9 … 18.4 μs) 3.92 μs ± 3.67% (3.58 … 4.53 μs) 20.0
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 219.0 μs ± 3.7% (205.0 … 230.0 μs) 187.0 μs ± 2.63% (179.0 … 194.0 μs) 87.0
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 230.0 μs ± 1.06% (224.0 … 237.0 μs) 185.0 μs ± 2.25% (175.0 … 190.0 μs) 78.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 32, 1), base shape (2, 16, 1) 21.9 μs ± 0.772% (21.7 … 22.2 μs) 4.31 μs ± 3.4% (4.05 … 4.53 μs) 18.7
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.46% (11.9 … 12.4 μs) 3.08 μs ± 2.98% (2.86 … 3.34 μs) 24.0
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 130.0 μs ± 0.22% (129.0 … 131.0 μs) 28.6 μs ± 0.532% (28.4 … 28.8 μs) 22.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (16, 16, 1), base shape (8, 4, 1) 14.5 μs ± 1.07% (14.3 … 14.8 μs) 4.09 μs ± 2.61% (3.81 … 4.53 μs) 26.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 8, 2), base shape (4, 8, 1) 12.4 μs ± 1.4% (12.2 … 12.6 μs) 4.04 μs ± 3.42% (3.81 … 4.29 μs) 31.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (16, 16, 1), base shape (4, 8, 1) 15.4 μs ± 1.11% (15.3 … 15.7 μs) 4.04 μs ± 3.41% (3.81 … 4.29 μs) 25.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.0 μs ± 1.35% (11.7 … 12.4 μs) 5.21 μs ± 3.34% (5.01 … 5.72 μs) 42.9
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.4 μs ± 1.54% (11.9 … 12.9 μs) 7.26 μs ± 1.94% (7.15 … 7.63 μs) 60.0
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.68 μs ± 2.14% (8.34 … 9.06 μs) 4.97 μs ± 3.42% (4.77 … 5.25 μs) 57.1
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 148.0 μs ± 0.162% (147.0 … 149.0 μs) 28.7 μs ± 0.488% (28.4 … 29.1 μs) 19.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.4 μs ± 1.33% (12.2 … 12.6 μs) 3.8 μs ± 3.24% (3.58 … 4.29 μs) 29.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.2 μs ± 1.33% (11.9 … 12.4 μs) 3.96 μs ± 4.04% (3.58 … 4.29 μs) 30.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 32, 1), base shape (2, 16, 1) 21.9 μs ± 0.737% (21.7 … 22.2 μs) 4.04 μs ± 3.57% (3.81 … 4.29 μs) 17.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (32, 4, 1), base shape (16, 2, 1) 19.1 μs ± 0.851% (18.8 … 19.3 μs) 4.04 μs ± 3.45% (3.81 … 4.29 μs) 20.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 8, 2), base shape (4, 8, 1) 11.9 μs ± 1.38% (11.7 … 12.2 μs) 4.3 μs ± 3.4% (4.05 … 4.53 μs) 34.7
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.2 μs ± 1.48% (11.9 … 12.4 μs) 5.21 μs ± 3.26% (5.01 … 5.48 μs) 42.0
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.67 μs ± 2.01% (8.34 … 9.06 μs) 5.11 μs ± 3.01% (4.77 … 5.48 μs) 57.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 16, 1), base shape (4, 8, 1) 15.2 μs ± 1.11% (15.0 … 15.5 μs) 4.31 μs ± 3.35% (4.05 … 4.53 μs) 27.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.2 μs ± 1.21% (11.9 … 12.4 μs) 3.94 μs ± 3.8% (3.81 … 4.29 μs) 32.0
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 46.4 ms ± 0.616% (46.4 … 52.8 ms) 40.8 ms ± 0.00621% (40.8 … 40.8 ms) 88.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (8, 8, 1), base shape (4, 8, 1) 15.4 μs ± 0.826% (15.0 … 15.7 μs) 4.09 μs ± 2.39% (3.81 … 4.29 μs) 25.4
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.82 μs ± 2.1% (8.34 … 9.06 μs) 4.95 μs ± 3.41% (4.77 … 5.25 μs) 57.1
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.4 μs ± 1.4% (11.9 … 12.6 μs) 7.37 μs ± 1.7% (7.15 … 7.87 μs) 60.0
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 377.0 μs ± 4.89% (349.0 … 396.0 μs) 293.0 μs ± 0.289% (290.0 … 294.0 μs) 83.3
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.3 μs ± 1.44% (11.9 … 12.6 μs) 7.29 μs ± 2.02% (7.15 … 7.63 μs) 60.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 8, 1), base shape (4, 8, 1) 13.4 μs ± 1.14% (13.1 … 13.8 μs) 4.1 μs ± 2.6% (3.81 … 4.53 μs) 29.1
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 9.23 μs ± 1.5% (8.82 … 9.54 μs) 2.21 μs ± 8.25% (1.67 … 2.62 μs) 18.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 9.33 μs ± 1.82% (9.06 … 9.54 μs) 2.93 μs ± 5.24% (2.62 … 3.1 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 8.0 μs ± 2.01% (7.87 … 8.34 μs) 2.04 μs ± 6.05% (1.67 … 2.38 μs) 21.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (16, 16, 1), base shape (8, 4, 1) 14.6 μs ± 0.895% (14.3 … 15.0 μs) 4.21 μs ± 3.89% (4.05 … 4.77 μs) 28.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 32, 1), base shape (1, 32, 1) 33.0 μs ± 0.422% (32.7 … 33.4 μs) 4.03 μs ± 3.5% (3.81 … 4.29 μs) 11.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 16, 1), base shape (4, 8, 1) 15.2 μs ± 1.14% (15.0 … 15.5 μs) 4.05 μs ± 3.42% (3.81 … 4.29 μs) 25.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (32, 4, 1), base shape (16, 2, 1) 19.2 μs ± 0.798% (18.8 … 19.3 μs) 4.19 μs ± 3.73% (4.05 … 4.53 μs) 21.5
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 2.37 ms ± 3.43% (2.32 … 2.63 ms) 867.0 μs ± 0.868% (823.0 … 877.0 μs) 35.4
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 282.0 μs ± 5.85% (256.0 … 295.0 μs) 109.0 μs ± 0.156% (109.0 … 110.0 μs) 42.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (8, 8, 1), base shape (4, 8, 1) 15.3 μs ± 1.0% (15.0 … 15.5 μs) 4.04 μs ± 3.47% (3.81 … 4.29 μs) 25.4
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 7.46 μs ± 1.73% (7.15 … 7.87 μs) 3.64 μs ± 3.18% (3.58 … 4.05 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.2 μs ± 1.42% (11.9 … 12.4 μs) 2.99 μs ± 4.76% (2.62 … 3.1 μs) 22.0
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 8.17 μs ± 1.68% (7.87 … 8.58 μs) 2.17 μs ± 7.97% (1.91 … 2.38 μs) 24.2
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.5 μs ± 1.44% (12.2 … 12.9 μs) 7.34 μs ± 1.94% (7.15 … 7.63 μs) 58.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.4 μs ± 1.41% (12.2 … 12.6 μs) 3.92 μs ± 3.77% (3.58 … 4.29 μs) 29.4
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 15.3 μs ± 1.09% (15.0 … 15.5 μs) 6.8 μs ± 2.04% (6.44 … 6.91 μs) 42.9
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (8, 8, 1), base shape (4, 8, 1) 15.4 μs ± 1.06% (15.3 … 15.7 μs) 4.19 μs ± 3.72% (4.05 … 4.53 μs) 26.6
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 287.0 μs ± 4.26% (256.0 … 293.0 μs) 109.0 μs ± 0.155% (109.0 … 109.0 μs) 42.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 16, 1), base shape (4, 8, 1) 15.3 μs ± 0.943% (15.0 … 15.5 μs) 4.19 μs ± 3.67% (4.05 … 4.53 μs) 27.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 9.19 μs ± 1.5% (8.82 … 9.54 μs) 2.9 μs ± 4.96% (2.62 … 3.1 μs) 29.7
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 9.21 μs ± 1.32% (8.82 … 9.54 μs) 2.22 μs ± 7.74% (1.91 … 2.38 μs) 21.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 8, 1), base shape (4, 8, 1) 15.2 μs ± 1.13% (15.0 … 15.5 μs) 4.3 μs ± 3.32% (4.05 … 4.77 μs) 27.0
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 15.8 μs ± 1.08% (15.5 … 16.0 μs) 3.95 μs ± 4.08% (3.58 … 4.29 μs) 23.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (16, 16, 1), base shape (4, 8, 1) 15.3 μs ± 0.921% (15.0 … 15.5 μs) 4.19 μs ± 3.77% (4.05 … 4.53 μs) 27.0
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.4 μs ± 1.37% (12.2 … 12.6 μs) 2.77 μs ± 4.99% (2.38 … 2.86 μs) 19.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 8, 2), base shape (4, 8, 1) 14.9 μs ± 1.08% (14.5 … 15.3 μs) 4.09 μs ± 2.43% (3.81 … 4.53 μs) 26.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (32, 4, 1), base shape (32, 1, 1) 31.6 μs ± 0.459% (31.5 … 31.9 μs) 4.3 μs ± 3.27% (4.05 … 4.53 μs) 12.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 9.18 μs ± 1.56% (8.82 … 9.54 μs) 2.91 μs ± 5.18% (2.62 … 3.1 μs) 29.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 32, 1), base shape (2, 16, 1) 21.9 μs ± 0.75% (21.7 … 22.2 μs) 4.19 μs ± 3.81% (4.05 … 4.53 μs) 18.7
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 15.7 μs ± 1.1% (15.5 … 16.0 μs) 3.96 μs ± 3.89% (3.81 … 4.29 μs) 24.6
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 1.55 ms ± 3.1% (1.52 … 1.68 ms) 293.0 μs ± 0.301% (290.0 … 294.0 μs) 19.1
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 46.3 ms ± 0.651% (46.3 … 52.7 ms) 40.8 ms ± 0.00609% (40.8 … 40.8 ms) 88.2
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.39% (11.9 … 12.4 μs) 5.08 μs ± 3.42% (4.77 … 5.72 μs) 40.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 32, 1), base shape (2, 16, 1) 22.0 μs ± 0.604% (21.7 … 22.2 μs) 4.09 μs ± 2.52% (3.81 … 4.53 μs) 17.6
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 12.0 μs ± 1.34% (11.7 … 12.4 μs) 4.51 μs ± 3.76% (4.29 … 4.77 μs) 36.7
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 18.2 μs ± 0.79% (17.9 … 18.6 μs) 6.76 μs ± 2.1% (6.44 … 6.91 μs) 36.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 16, 1), base shape (4, 8, 1) 15.2 μs ± 1.14% (15.0 … 15.5 μs) 4.09 μs ± 2.45% (3.81 … 4.29 μs) 25.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 8, 1), base shape (4, 8, 1) 15.3 μs ± 1.1% (15.0 … 15.5 μs) 4.19 μs ± 3.74% (4.05 … 4.53 μs) 27.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 32, 1), base shape (1, 32, 1) 33.0 μs ± 0.438% (32.7 … 33.4 μs) 4.3 μs ± 3.42% (4.05 … 4.53 μs) 12.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 12.1 μs ± 1.2% (11.7 … 12.4 μs) 4.31 μs ± 3.27% (4.05 … 4.53 μs) 34.7
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 138.0 μs ± 6.53% (129.0 … 150.0 μs) 32.6 μs ± 0.508% (32.4 … 32.9 μs) 25.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (16, 16, 1), base shape (4, 8, 1) 15.2 μs ± 1.09% (15.0 … 15.5 μs) 4.3 μs ± 3.34% (4.05 … 4.77 μs) 27.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 7.43 μs ± 1.48% (7.15 … 7.87 μs) 3.58 μs ± 2.83% (3.1 … 4.05 μs) 43.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (16, 16, 1), base shape (8, 4, 1) 14.5 μs ± 1.09% (14.3 … 14.8 μs) 4.3 μs ± 3.33% (4.05 … 4.77 μs) 28.3
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 12.6 μs ± 1.36% (12.4 … 12.9 μs) 7.16 μs ± 1.26% (6.91 … 7.63 μs) 55.8

@codecov Codecov
Copy link

codecov bot commented Nov 14, 2023

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (8a73476) 33.58% compared to head (57ee67e) 34.02%.

Additional details and impacted files 
@@            Coverage Diff             @@
##           master     #172      +/-   ##
==========================================
+ Coverage   33.58%   34.02%   +0.43%     
==========================================
  Files          11       11              
  Lines         914      920       +6     
==========================================
+ Hits          307      313       +6     
  Misses        607      607

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

@thomasfaingnaert thomasfaingnaert changed the title Fix occasional incorrect default value for memory Improve heuristic for memcopy tile sizes Nov 14, 2023
@thomasfaingnaert thomasfaingnaert merged commit 433aa68 into master Nov 14, 2023
1 check passed
@thomasfaingnaert thomasfaingnaert deleted the tf/fix-memory-heuristic branch November 14, 2023 18:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@thomasfaingnaert @maleadt