From 5ff0ed390ac566a369d1f718273c99f6a13045eb Mon Sep 17 00:00:00 2001 From: papa Date: Sat, 14 Nov 2026 20:58:26 +0100 Subject: [PATCH] Add best practice: Real-time 2D sprite blending and alpha modulation with AMMX --- .../fast-alpha-blitting-pmula.md | 69 +++++++++++++++++++ 1 file changed, 69 insertions(+) create mode 100644 core/best-practices/fast-alpha-blitting-pmula.md diff --git a/core/best-practices/fast-alpha-blitting-pmula.md b/core/best-practices/fast-alpha-blitting-pmula.md new file mode 100644 index 0000000..25b1f79 --- /dev/null +++ b/core/best-practices/fast-alpha-blitting-pmula.md @@ -0,0 +1,69 @@ +--- +title: Real-time 2D sprite blending and alpha modulation with AMMX +type: best-practice +owner: Rudi7503 +status: draft +origin: human +basis: [measured] +components: [cpu-68080, saga-video-isabelle] +applies_to: + core: "any" + os: [any] +tags: [68080, ammx, blit, sprite, alpha, blending, games, hurrican] +related: + - core/best-practices/ammx-where-it-wins.md + - core/best-practices/align-copy-destination.md +sources: + - "Hardware measurement on Vampire V4 Standalone (Core 11350 at 92 MHz), Harness v6 suite (BLT-006..021)" +--- + +## Goal + +Achieve high frame-rate 2D sprite rendering, alpha blending, dynamic lighting, and particle effects in games without frame drops or blitter stalls. + +## Technique + +- **Vectorized Color-Key Transparency (`BLT-013`):** + - Compare 8 pixels (or 4 words) in parallel using `pcmpccw` / `pcmpccb`. + - Use the resulting bitmask to conditionally mask pixel writes with `storem` or `storeilm`. +- **Fast Alpha Blending via `pmula` (`BLT-006`):** + - Use `pmula` (multiply-accumulate with alpha scaling) to blend source and destination color channels in parallel. +- **Fast Additive/Particle Blending (`BLT-018`):** + - For light, glow, and fire particles, use saturating addition `paddusw` / `paddusb` across 64-bit qwords. + - Saturated SIMD math prevents color wrapping artifacts without branch checks. +- **Branchless Row Tails (`BLT-012`):** + - Terminate sprite rows with `storeilm` using a preloaded lane mask instead of branching to scalar cleanup loops. + +```asm +; kb:compile + section code,code + ; Example: Fast 64-bit additive particle blend step + load (a0),e0 ; Load 4 source pixels (RGB565 or 8-bit) + load (a1),e1 ; Load 4 destination pixels + paddusw e0,e1,e2 ; Saturated vector add (no color wrap) + store e2,(a1) ; Write back blended pixels +``` + +## Evidence + +Measured on physical Vampire V4 Standalone hardware (Core 11350 at 92 MHz, Harness v6 suite, 10-iteration loop baseline 3.0 cycles): + +| Blit / Blend Operation | Kernel ID | Net Cycles / Iteration | Fill Rate (92 MHz) | Architectural Feature | +|---|:---:|:---:|:---:|---| +| **Ultra-Fast Qword Blend** | `BLT-018` | **0.6 cyc** | **1,226.7 Mpx/s** | Saturating SIMD add, > 1.2 GPx/s throughput | +| **Color-Key Transparency Filter** | `BLT-013` | **1.6 cyc** | **460.0 Mpx/s** | 1-cycle parallel color key match for 4/8 pixels | +| **Alpha Blit Basic Step** | `BLT-006` | **1.6 cyc** | **460.0 Mpx/s** | Parallel channel modulation | +| **Masked Qword Blit** | `BLT-012` | **2.6 cyc** | **283.1 Mpx/s** | Tail masking via `storeilm` | +| **Subtractive Blend (Shadows)** | `BLT-014` | **4.6 cyc** | **160.0 Mpx/s** | Saturating vector subtraction `psubusw` | +| **Multiplicative Blend (Darkening)**| `BLT-015` | **4.6 cyc** | **160.0 Mpx/s** | Fixed-point darkening | +| **Particle Shading (8 Pixels)** | `BLT-016` | **12.6 cyc** | **58.4 Mpx/s** | 8 particles rendered simultaneously | + +### Key Engine Insights + +- **Eliminating Blitter Wait States**: Unlike the classic Amiga blitter which halts CPU execution or requires busy-wait polling (`btst #6,$DFF002`), CPU-driven AMMX blits execute asynchronously out of L1 cache and stream directly to DDR3 video memory. +- **Saturating Math**: `paddusw` and `psubusw` provide hardware clipping at `$FFFF` (white) and `$0000` (black) with zero branch penalties. + +## When not to use + +- When using the SAGA 3D hardware rasterizer (Maggie): if Maggie spans and textured polygons are already active, let Maggie handle depth-tested surfaces instead of software AMMX spans. +- For small 1-to-2 pixel wide vertical blits: scalar 16-bit stores avoid SIMD masking overhead.