Skip to content

Latest commit

 

History

History
327 lines (257 loc) · 11.7 KB

PowerVR_BXM.md

File metadata and controls

327 lines (257 loc) · 11.7 KB

Imagination Technologies PowerVR BXM-8-256

Specs

  • Execution units: 8
  • Warp width: 16
  • Vulkan subgroup size: 128 (= Total ALUs = EU * warp)
  • Total ALUs: 128 - number of simultaneously executing threads
  • Clock: 950 MHz
  • F16 GFLOPS: 460 (180 on MulAdd from tests)
  • F32 GFLOPS: 230 (190 on FMA from tests)
  • FP16 FLOPs/Clock: 512
  • FP32 FLOPs/Clock: 256 - 2FLOPS per clock for FMA
  • Memory: 8 GB, LPDDR5, QC 16bit, 3200 MHz, 51.2 GB/s (14.2 GB/s from tests)
  • Device: Motorola G54 5G (MediaTek Dimensity 7020, Android 13, Driver 6133109)

Theoretical performance:

FLOPS = clock * warp_width * EU
FLOPs/Clock = warp_width * EU * 2 (FMA == 2FLOPS)

950M * 128 = 121.6G FMA ops per second = 243.2 GFLOPS

Shader

Quads

  • Quads on edge between 2 triangles are not merged, so 2 near pixels may execute up to 6 helper invocations.
  • Test subgroupQuadBroadcast( gl_HelperInvocation ) with/without texturing - helper invocations are executed. [6]
  • Test subgroupQuadBroadcast( constant ) with/without texturing - helper invocations are executed. [6]

Subgroups

  • Subgroups in fragment shader can fill multiple triangles, but only with the same gl_InstanceIndex. [6]
  • Subgroups in fragment shader always execute all threads. It cause a 128 threads to be executed which is bad for energy efficiency. [6]

Subgroup threads order

Result of Rainbow( gl_SubgroupInvocationID / gl_SubgroupSize ) in fragment shader, gl_SubgroupSize: 128, image size: 32x32. [6]

Result of Rainbow( gl_SubgroupInvocationID / gl_SubgroupSize ) in compute shader, gl_SubgroupSize: 128, workgroup size: 8x8. [6]

Result of Rainbow( gl_SubgroupInvocationID / gl_SubgroupSize ) in compute shader, gl_SubgroupSize: 128, workgroup size: 16x16. [6]

Instruction cost

  • Shader instruction benchmark results: [4]

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

    • fp32 Pow uses MUL loop - performance depends on power

    • fp16 is a bit slower than fp32 because of conversion between fp16 and fp32

    • base rate: 128 GOp/s

    • float point

    op \ type fp16 fp32
    Add 1 1
    Mul 1 1
    FMA 1 1
    MulAdd 1 1
    Lerp 2 2
    Length 2 2
    Normalize 2 2
    Distance 2.5 2.5
    Dot 3 3
    Cross 3 3
    Min/Max 2 2
    Clamp(x,0,1) 0-1 0-1
    Clamp(x,-1,1) 3 3
    Clamp 4 4
    Step 2 1.5
    SmoothStep 4 1.5
    Abs 0-1 0-1
    SignOrZero 4 2.5
    BitCast 2 2
    FloatToInt 4 3
    IntToFloat 4 3
    Ceil, Floor, Fract 2 1
    Trunc 9 9
    Round, RoundEven 10 10
    Exp, Exp2 10 10
    Log, Log2 10 10
    InvSqrt 2 2
    Sqrt 5 4.5
    Sin, Cos 32 32
    Div 4 2
    Mod 5 5
    Pow 2-20 2-20
    Tan 32 32
    ASin, ACos 20-32 20-32
    ATan 67 67
    • float point fast math

    | op \ type | fp16 | fp32 | |---|---|---|---| | fast Sign | 3 | 1 | | Cbrt (pow) | 21 | 21 | | Cbrt (exp) | 21 | 21 | | sRGB | 22 | 22 | | fast sRGB | 7 | 7 | | fast Cos | | fast Sin | | fast Tan | | fast ASin | 8 | 9 | | fast ACos | 9 | 9 | | fast ATan | 28 | 22 | | fast ASin v2 | | fast ACos v2 | | fast ATan v2 |

    • integer
    op \ type i32 u32 i16 u16 u8 i64 u64
    Add 3 3 3 3 6 4
    Mul 3 3 3 3 3 12 10
    MulAdd 3 3 3 3 18 14
    Div 120 110 14 14 1420 268
    Mod 120 110 28 14 2520 2060
    Min/Max 3 3 3 3 9 9
    Clamp const 6 3 6 3 18 10
    Clamp 3 3 3 3 6 12 11
    Abs 1 - 3 - 6 -
    SignOrZero
    Shift const 1 1 1 0.5 6 22
    Shift 36 17 3 3 100 22
    And 3 3 3 3 6 6
    Or 3 3 3 3 6 6
    Xor 3 3 3 3 6 6
    BitCount 3 3 - - - -
    FindLSB 7 7 10 10 10 14 12
    FindMSB 16 3 - - - - -
    AddCarry - 10 - - - - -
    SubBorrow - 11 - - - - -
    MulExtended 6 7 - - - - -

  • FP32 instruction performance: [2]

    • Loop unrolling doesn't increase performance.
    • Loop unrolling is too slow at pipeline creation stage.
    • Manual unrolling is slow too and performance is less than with unrolling attribute.
    • Compute and graphics has same performance.
    • Dispatch on 1024x1024 grid is much faster (1.3x).
    • Loop index with int is faster than float.
    • mediump has no effect.
    • Measured at 950 MHz with 87% shader load.
    GOp/s ops max GFLOPS
    105 Add 105
    95 Mul 95
    95 MulAdd, FMA 190

  • FP16 instruction performance: [1]

    • Loop index with int, short, half has same performance.
    • Measured at 950 MHz
    GOp/s ops max GFLOPS comments
    110 Add 110
    90 Mul 90
    90 MulAdd 180
    58 FMA 116 less than F32FMA

NaN / Inf

  • FP32, Mediump. [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 0 0 0 0 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) nan nan nan nan 1 -1 1 -1
    SignOrZero(-x) nan nan nan nan -1 1 -1 1
    SmoothStep(x,0,1) 0 0 0 0 1 0 1 0
    Normalize(x) nan nan nan nan nan nan 18446742974197923840 -18446742974197923840

  • FP16 diff:

    op \ type nan1 nan2 nan3 nan4 inf -inf max -max
    Normalize(x) nan nan nan nan nan nan 0 -0

Shared memory

  • External traffic is too low - shared memory is used. [10]

Noise performance

Circle performance

  • small circles. [13]

    • 32K objects
    • 41.4 MPix
    • driver drop render passes which will be overriden by next render pass !!! keep only 3 RPs
    shape exec time (ms) diff (%)
    quad 11.1 -
    fan 14.9
    strip 15.5
    max area 14.3

  • 4x4 circles with blending. [13]

    • 10.4 MPix
    • 64 layers
    shape exec time (ms) diff (%)
    quad 83.5 -
    fan 65.9 26.7
    strip 65.3 27.9
    max area 65.9 26.7

Branching

  • Mul vs Branch vs Matrix [12]

    • 1.05 MPix, 128 iter, 6 mul/branch ops.
    op exec time (ms) diff
    Mul uniform 50.8 2.3
    Branch uniform 22.3 -
    Matrix uniform 34.4 1.5
    -
    Mul non-uniform 59.0 2.6
    Branch non-uniform 78.1 3.5
    Matrix non-uniform 69.5 3.1
    -
    Mul avg 54.9 2.46
    Branch avg 50.2 2.25
    Matrix avg 51.9 2.33

Resource access

  • Buffer/Image storage 16bpp 2.59MPix 2x41.4MB [7]

    diff exec time (ms) approx traffic (GB/s) name comments
    1.09 6.18 13.2 Image load/store
    1 5.6 14.2 Image read/write input attachment RGBA32F a bit faster because of RT compression
    Image read/write input attachment 4xRGBA8
    2.7 15 3.8 Buffer load/store ???
    3.2 18 3.3 Buffer load/store in FS

Render target compression

  • RGBA8 67.1MPix downsample 1/2, compressed/uncompressed access rate: [3]

    • expected read: 268MB, write: 67MB, total: 335MB per frame.
    • with solid color: linear: 19.4ms, fetch: 17.7ms, nearest: 17.7ms. Fetch/Nearest minimize bus load.
    • with gradient: linear/fetch/nearest has same perf.
    • graphics to compute r/w: 268MB / 66MB. Compression disabled when used storage usage flag.
    diff (read) read (MB) write (MB) name comments
    1 268 66 image storage
    1.33 202 50 1x1 noise
    1.35 198 51 2x2 noise
    2.4 112 50 4x4 noise
    13 21 7 gradient
    23 11.5 27 8x8 noise same as block size
    23 11.5 3.5 16x16 noise less write traffic because output to 8x8 block
    134 2 1 solid color has metadata for large region or small metadata for block

  • RGBA16F 67.1MPix downsample 1/2, compressed/uncompressed access rate: [3]

    • expected read: 536.8MB, write: 134.2MB, total: 671MB per frame.
    • graphics to compute r/w: 530MB / 130MB. Compression disabled when used storage usage flag.
    • image storage read with linear filter: 45ms, nearest/load: 61.
    • 1x1 noise gradient: linear: 35ms, fetch/nearest: 50ms.
    diff (read) read (MB) write (MB) name comments
    1 530 130 image storage
    1.3 410 90 1x1 noise
    1.4 390 105 2x2 noise
    2.6 205 92 4x4 noise
    9.8 55 17 gradient
    24 22 55 8x8 noise same as block size
    27 20 5.5 16x16 noise less write traffic because output to 8x8 block
    134 4 1.5 solid color has metadata for large region or small metadata for block

  • RGBA16_UNorm - same as RGBA16F.

  • RGBA32F - has compression, but without linear filtering.

Texture cache

  • RGBA8_UNorm texture with random access [9]

    • Measured cache size: 256 KB, 1 MB.
    size (KB) dimension (px) external bandwidth (GB/s) comment
    256 256x256 0.009 used only texture cache
    1024 512x512 13.9 bottleneck on external memory