Skip to content

Latest commit

 

History

History
346 lines (278 loc) · 13.9 KB

Adreno_660.md

File metadata and controls

346 lines (278 loc) · 13.9 KB

Qualcomm Adreno 660

Specs

  • Clock: 790 MHz
  • Cores: 2
  • Total ALUs: 1024 - number of simultaneously executing threads
  • F16 GFLOPS: 3244 (1414 on MulAdd from tests)
  • F32 GFLOPS: 1622 (728 on FMA from tests)
  • F64 GFLOPS: 405
  • GMem size: 1.5 Mb (bandwidth?)
  • L2: ? (bandwidth?)
  • Memory: 8 GB, LPDDR5-6400, QC 16bit, 3200MHz, 51.2 GB/s (35 GB/s from tests)
  • Device: Asus ROG Phone 5 (Android 13, Driver 512.530.0)

Theoretical performance:

FLOPS = clock * TotalALUs

790M * 1024 = 809G FMA ops per second = 1620 GFLOPS

Shader

Quads

  • Test subgroupQuadBroadcast( gl_HelperInvocation ) without texturing - helper invocations are not executed (or not detected). [6]

  • Test subgroupQuadBroadcast( gl_HelperInvocation ) with texturing - helper invocations are executed, even if Nearest immutable sampler is used. [6]
    Red - no helper invocations, violet - 3 helper invocations per quad.

  • Test subgroupQuadBroadcast( constant ) with/without texturing - helper invocations are executed. [6]

Subgroups

  • Subgroups in fragment shader can fill multiple triangles, even if they have different gl_InstanceIndex. [6]

  • Subgroups in fragment shader reserve threads for helper invocations, even if they are not executed. [6]

  • Subgroup occupancy for single triangle with texturing. Helper invocations are executed and included as active thread. Red color - full subgroup. [6]

  • Subgroup occupancy for single triangle without texturing. Helper invocations are not executed but threads are reserved, so occupancy is low. Red color - full subgroup. [6]

  • Subgroup occupancy for too small triangles. Red color - full subgroup. [6]

Subgroup threads order

Result of Rainbow( gl_SubgroupInvocationID / gl_SubgroupSize ) in fragment shader, gl_SubgroupSize: 64. [6]

Result of Rainbow( gl_SubgroupInvocationID / gl_SubgroupSize ) in compute shader, gl_SubgroupSize: 64, workgroup size: 8x8. [6]

Instruction cost

  • Shader instruction benchmark results: [4]

    • fp32 & i32 datapaths can execute in parallel in 2:1 rate.

    • i8 has same performance as i16, except i8 Mod which is 2x slower.

    • u8 has same performance as u16.

    • int/uint mediump/lowp has same performance as i16/u16.

    • mediump BitCount is 5x faster than i32.

    • base rate: 420 GOp/s

    • float point

    op \ type fp16 fp32
    Add 0.5 1
    Mul 0.5 1
    FMA 3 1-1.5
    MulAdd 0.5 1-1.5
    Lerp 1.5 3
    Length 1.5 1
    Normalize 1 1.5
    Distance 1.5 2
    Dot 1 2
    Cross 1.5 2
    Min/Max 0.5 1
    Clamp(x,0,1) 1 2
    Clamp(x,-1,1) 1 2
    Clamp 1 2
    Step 1.5 2
    SmoothStep 1 3
    Abs 0.5
    FastSign 3 4
    SignOrZero 0.5 1
    BitCast 0.5 1
    FloatToInt 1 1
    IntToFloat 0.5 1
    Ceil, Floor, Trunc, Round, RoundEven 1 1
    Fract 1 1.5
    Exp, Exp2 3 6
    Log, Log2 3 6
    InvSqrt 3 6
    Sqrt 6 6
    Sin, Cos 6 6
    Div 6 6
    Mod 6 6
    Pow 1-6 1.5-14
    Tan 15 24
    ASin, ACos 150 160
    ATan 100 120
    • integer
    op \ type i32 u32 i16 u16
    Add 1 1 0.5 0.5
    Mul 4 4 0.5 0.5
    MulAdd 5 5 1.5 1
    Div 42 24 14 7
    Mod 64 28 16 9
    Min/Max 1 1 0.5 0.5
    Clamp const 2 1 1 0.5
    Clamp 1 1 0.5 0.5
    Abs 1 - 0.5 -
    SignOrZero
    Shift const 1 1 0.5 0.5
    Shift 1 1 1 1
    And 1 1 0.5 0.5
    Or 1 1 0.5 0.5
    Xor 1 1 0.5 0.5
    BitCount 6 6 - -
    FindLSB 2 2 1.5 1.5
    FindMSB 4 3 - -
    AddCarry - 4 - -
    SubBorrow - 4 - -
    MulExtended 20 20 - -

  • FP32 instruction performance: [2]

    • Loop unrolling is fast during pipeline creation if loop < 256.
    • Loop unrolling is 1x - 1.4x faster, 2x slower on 1024, 1.1x slower on 256.
    • Loop index with int and float has same performance.
    • Compute dispatch on 2Kx2K grid is faster, 256x256 grid is a bit slower.
    • Graphics render area 2Kx2K is faster.
    • 128 subgroup size (wave128) has no effect on performance.
    GOp/s exec time (ms) ops max GFLOPS
    420 10.2 Add, Mul 420
    364 11.8 FMA, MulAdd 728

  • FP16 instruction performance: [1]

    GOp/s exec time (ms) ops max GFLOPS
    830 5.16 Add, Mul 830
    707 6.06 MulAdd 1414
    117 36.5 FMA 234

NaN / Inf

  • FP32, FP16. [11]

    op \ type nan1 nan2 nan3 nan4 inf -inf max -max
    x nan nan nan nan inf -inf max -max
    Min(x,0) 0 0 0 0 0 -inf 0 -max
    Min(0,x) 0 0 0 0 0 -inf 0 -max
    Max(x,0) 0 0 0 0 inf 0 max 0
    Max(0,x) 0 0 0 0 inf 0 max 0
    Clamp(x,0,1) 0 0 0 0 1 0 1 0
    Clamp(x,-1,1) -1 -1 -1 -1 1 -1 1 -1
    IsNaN 1 1 1 1 0 0 0 0
    IsInfinity 0 0 0 0 1 1 0 0
    bool(x) 1 1 1 1 1 1 1 1
    x != x 1 1 1 1 0 0 0 0
    Step(0,x) 1 1 1 1 1 0 1 0
    Step(x,0) 1 1 1 1 0 1 0 1
    Step(0,-x) 1 1 1 1 0 1 0 1
    Step(-x,0) 1 1 1 1 1 0 1 0
    SignOrZero(x) 0 0 0 0 1 -1 1 -1
    SignOrZero(-x) 0 0 0 0 -1 1 -1 1
    SmoothStep(x,0,1) 0 0 0 0 1 0 1 0
    Normalize(x) nan nan nan nan nan nan 0 -0

  • FP Mediump diff:

    op \ type nan1 nan2 nan3 nan4 inf -inf max -max
    x nan nan nan nan inf -inf 65504 -65504
    Min(x,0) 0 0 0 0 0 -inf 0 -65504
    Min(0,x) 0 0 0 0 0 -inf 0 -65504
    Max(x,0) 0 0 0 0 inf 0 65504 0
    Max(0,x) 0 0 0 0 inf 0 65504 0

Shared memory

TODO

Noise performance

Circle performance

  • small circles. [13]

    • 8K objects
    • 10.5MPix
    shape exec time (ms) diff (%)
    quad 2.7 -
    fan 3.4 26
    strip 3.3 22
    max area 3.3 22

  • 4x4 circles with blending. [13]

    • 10.5MPix
    • 64 layers
    shape exec time (ms) diff (%)
    quad 69.5 -
    fan 40.6 71
    strip 40.4 72
    max area 40.5 71

Branching

  • Mul vs Branch vs Matrix [12]

    • 4.2 MPix, 128 iter, 6 mul/branch ops.
    op exec time (ms) diff
    Mul uniform 49.0 1.6
    Branch uniform 30.7 -
    Matrix uniform 30.8 1.0
    -
    Mul non-uniform 70.8 2.3
    Branch non-uniform 56.6 1.8
    Matrix non-uniform 92.1 3.0
    -
    Mul avg 60.0 1.95
    Branch avg 43.7 1.4
    Matrix avg 61.4 2.0

Blending

TODO

Resource access

  • Texture access 67.1MPix: [5]

    • expected read: 268MB per frame.
    • UV bias has no effect.
    diff exec time (ms) approx traffic (GB/s) name comments
    0.5 5 53 sequential access, scale x0.5 too high calculated bandwidth because of texture cache
    1 9.6 28 sequential access, scale x1 near to external memory bandwidth
    1.3 12.6 21 random access, noise 16x16
    2 19 14 sequential access, scale x1.5
    2.1 20.5 13 random access, noise 8x8
    3.2 31 8.6 sequential access, scale x2
    5.4 52.5 5.1 random access, noise 4x4
    5.6 53.7 5 random access, noise 4x4, off 1
    15.6 150 1.78 random access, noise 2x2
    15.8 152 1.76 random access, noise 2x2, off 1
    60 580 0.46 random access, noise 1x1
    60 578 0.46 random access, noise 1x1, off 1

  • Buffer/Image storage 16bpp 16.7MPix 2x267MB [7]

    diff (%) exec time (ms) approx traffic (GB/s) name comments
    -33 11.4 47 Image read/write input attachment 2xRGBA16, noise 16x16 noise size equal to RT compressed block
    -28 11.9 45 Image read/write input attachment 2xRGBA16, noise 8x8
    -22 12.5 43 Image read/write input attachment 2xRGBA16, noise 4x4
    0 15.2 35 Buffer load/store, 16 byte, wg:64 16 byte cache line (?)
    2 15.5 34 Image load/store, wg 16x16, row major
    5 16.0 33 Image read/write input attachment 2xRGBA16, noise 2x2
    15 17.5 31 Buffer load/store, 64 byte, wg:64
    15 17.5 31 Image load/store, wg 8x8, row major
    15 17.5 31 Image load/store, wg 8x8, group reorder, row major group reorder has no effect
    18 18.0 30 Buffer load/store, 128 byte, wg:64
    20 18.3 29 Buffer load/store, 16 byte, wg:8x8
    25 19.0 28 Buffer load/store, 128 byte, wg:16x16
    26 19.2 28 Buffer load/store, 64 byte, wg:16x16
    27 19.3 28 Buffer load/store, 32 byte, wg:64
    33 20.2 26 Buffer load/store, 128 byte, wg:8x8
    Image fetch/sample in CS with double buffering
    49 22.6 24 Image fetch/sample in FS with double buffering a bit faster than input attachment because of texture cache
    53 23.3 23 Buffer load/store, 16 byte, wg:16x16
    53 23.3 23 Buffer load/store in FS
    55 23.5 23 Image read/write input attachment RGBA32F RT compression is not supported, same performance as buffer in FS - cache misses because of thread order(?), immediate mode rendering should be used
    55 23.5 23 Buffer load/store, 64 byte, wg:8x8
    74 26.4 20 Image load/store, wg 8x8, column major
    61 24.4 22 Buffer load/store, 32 byte, wg:16x16
    91 29.0 18 Buffer load/store, 32 byte, wg:8x8

Render target compression

  • RGBA8 268MPix downsample 1/2, compressed/uncompressed access: [3]

    • expected read: 1.07GB, write: 268MB, total: 1.34GB per frame.
    • 8x8 noise linear/fetch/nearest has same perf.
    • 16x16 noise linear is faster. Linear filter minimize L2 cache misses on high compression rate.
    • graphics to compute: solid color: 68ms, gradient: 72ms, 1x1 noise: 80ms.
    • with storage usage flag - perf same as for graphics.
    diff exec time (ms) approx traffic (GB/s) name comments
    1 72 18.6 image storage
    1.7 42 31.9 1x1 noise near to external memory bandwidth
    1.6 44 30.4 2x2 noise
    2.2 32 41.9 4x4 noise
    2.5 29 46.2 8x8 noise
    3.3 22 60.9 16x16 noise same as block size
    3.4 21 63.8 gradient
    6.9 10.4 128.8 solid color

  • RGBA16_UNorm, RGBA16F - RT compression is supported. [3]

    diff exec time (ms) approx traffic (GB/s) name
    1 150 image storage
    2 75 1x1 noise
    16x16 noise
    3.3 45 solid color

  • RGBA32F - RT compression is not supported. [3]

Texture cache

  • RGBA8_UNorm texture with random access [9]

    • Measured cache size: 2 KB, 4 KB (?), 128 KB.
    • 8 texels per pixel, dim ???
    size (KB) dimension (px) exec time (ms) diff approx bandwidth (GB/s) comments
    1 16x16 TODO
    2 32x16 2.3 TODO L1 cache
    4 32x32 7 3
    16 64x64 12.4 1.8
    128 256x128 14 L2 cache
    256 256x256 44 3